Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

Day 1: Benchmarks We ran 1000+ LLM benchmarks on real consumer devices. Data includes single-device and multi-device clusters with Tokens-Per-Second (TPS) and Time-To-First-Token (TTFT). Setups tested: 3x M4 Mac Mini cluster, iPhone 15 + S24, RTX4090 & more.

EXO Labs

52,561 subscribers

255,162 Aufrufe • vor 1 Jahr •via X (Twitter)

Nachrichten & Politik Wissenschaft & Technologie Bildung

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

Today, we're shipping MLX support for TADA, our open-source text-to-speech model, which means the entire pipeline (LLM, flow-matching, and decoder) can now run locally on any Apple Silicon device. We're seeing a 45% reduction in memory usage and a 10x speed-up when using it quantized. With these improvements, you can use TADA on-device for OpenClaw or any personal chatbot. If you own a MacBook, Mac Mini, or Mac Studio, record a 10-second clip of any voice, type any text, and get high-quality, natural and expressive speech in real-time. Completely offline, completely free.

Today, we're shipping MLX support for TADA, our open-source text-to-speech model, which means the entire pipeline (LLM, flow-matching, and decoder) can now run locally on any Apple Silicon device. We're seeing a 45% reduction in memory usage and a 10x speed-up when using it quantized. With these improvements, you can use TADA on-device for OpenClaw or any personal chatbot. If you own a MacBook, Mac Mini, or Mac Studio, record a 10-second clip of any voice, type any text, and get high-quality, natural and expressive speech in real-time. Completely offline, completely free.

Hume AI

24,765 Aufrufe • vor 6 Monaten

CRYPTIC ALPHA V1 IS NOW LIVE After months of focused development and rigorous testing, Cryptic Alpha V1 is officially live on Android. This marks the first time quantum-safe messaging and private token swaps are delivered together in a single, consumer-grade mobile app protected by device-level encryption and built for real-world use. For the first time, privacy at this scale is accessible to everyday users. New integrations, features, and expansions are already underway as we continue building the future of private digital life.

CRYPTIC ALPHA V1 IS NOW LIVE After months of focused development and rigorous testing, Cryptic Alpha V1 is officially live on Android. This marks the first time quantum-safe messaging and private token swaps are delivered together in a single, consumer-grade mobile app protected by device-level encryption and built for real-world use. For the first time, privacy at this scale is accessible to everyday users. New integrations, features, and expansions are already underway as we continue building the future of private digital life.

Cryptic

21,380 Aufrufe • vor 9 Monaten

Create and warm 50 TikTok accounts in the US. This is where modern content infrastructure is heading. Not VPN-based provisioning. Not disposable or low-trust setups. Real, in-country accounts created on physical devices, with local signals and proper warming. Why this matters: You can generate thousands of posts per day. Getting them seen is the hard part. TikTok evaluates more than just IP. SIM, GPS, device fingerprint, and early behavior all play a role. That’s why most setups fail within days. The teams that win separate the stack: generate → render → distribute And treat distribution as infrastructure, not an afterthought. If you want to go deeper: Setup breakdown: Ultimate content factory playbook:

Create and warm 50 TikTok accounts in the US. This is where modern content infrastructure is heading. Not VPN-based provisioning. Not disposable or low-trust setups. Real, in-country accounts created on physical devices, with local signals and proper warming. Why this matters: You can generate thousands of posts per day. Getting them seen is the hard part. TikTok evaluates more than just IP. SIM, GPS, device fingerprint, and early behavior all play a role. That’s why most setups fail within days. The teams that win separate the stack: generate → render → distribute And treat distribution as infrastructure, not an afterthought. If you want to go deeper: Setup breakdown: Ultimate content factory playbook:

CODIFY

61,985 Aufrufe • vor 5 Monaten

Does LLM really need to be a helpful assistant all the time? No. If you want to simulate people, “perfectly helpful” could be the wrong objective. Meet OdysSim, a journey toward LLMs beyond assistants, as behavioral foundation models (10B tokens of real human behavior; 23 sim benchmarks, finally in one place. new open models: outperform or on par with GPT-5.5, Gemini 3.1, or Claude Opus 4.7 in many behavior-sim dimensions). Human behavior simulation is becoming essential. Agent evaluation needs realistic users before real users show up. Medical and classroom training need realistic patients and students. Social science needs synthetic participants at scale. But real people are not ideal assistants. Real patients panic or ignore good advice. Real students misunderstand. Real customers are vague, picky, impatient, or simply leave. Human behavior is messy, diverse, and often imperfect. Frontier LLMs are getting better at math, code, and long-horizon tasks. They are NOT getting better at simulating human behavior. If anything, they drift the other way: more assistant-ish, more homogeneous, fewer of the errors and quirks real humans show. This is no accident. The whole pipeline is built for helpfulness and task success, not behavioral realism. And you can't prompt your way out of that. So we rethink the recipe from scratch and release: 🧠 The OdysSim corpus: 21.4M real human interactions (~10B tokens) from 62 sources, every conversation retrofitted with social grounding (who is talking, and why) 📏 SOUL-Index: 23 human-behavior benchmarks unified into one suite across 5 axes 🤖 OSim-8B: open weights; tops more SOUL-Index benchmarks than any frontier model, acts more like a real user than any of them on τ-bench (nearly matching real humans in the reaction dimension), and writes far more human-like text along the way.

Does LLM really need to be a helpful assistant all the time? No. If you want to simulate people, “perfectly helpful” could be the wrong objective. Meet OdysSim, a journey toward LLMs beyond assistants, as behavioral foundation models (10B tokens of real human behavior; 23 sim benchmarks, finally in one place. new open models: outperform or on par with GPT-5.5, Gemini 3.1, or Claude Opus 4.7 in many behavior-sim dimensions). Human behavior simulation is becoming essential. Agent evaluation needs realistic users before real users show up. Medical and classroom training need realistic patients and students. Social science needs synthetic participants at scale. But real people are not ideal assistants. Real patients panic or ignore good advice. Real students misunderstand. Real customers are vague, picky, impatient, or simply leave. Human behavior is messy, diverse, and often imperfect. Frontier LLMs are getting better at math, code, and long-horizon tasks. They are NOT getting better at simulating human behavior. If anything, they drift the other way: more assistant-ish, more homogeneous, fewer of the errors and quirks real humans show. This is no accident. The whole pipeline is built for helpfulness and task success, not behavioral realism. And you can't prompt your way out of that. So we rethink the recipe from scratch and release: 🧠 The OdysSim corpus: 21.4M real human interactions (~10B tokens) from 62 sources, every conversation retrofitted with social grounding (who is talking, and why) 📏 SOUL-Index: 23 human-behavior benchmarks unified into one suite across 5 axes 🤖 OSim-8B: open weights; tops more SOUL-Index benchmarks than any frontier model, acts more like a real user than any of them on τ-bench (nearly matching real humans in the reaction dimension), and writes far more human-like text along the way.

Xuhui Zhou

143,753 Aufrufe • vor 4 Monaten

AN AWS ENGINEER QUIETLY BUILT A 2 PETABYTE HOME SERVER FOR $9/MONTH THAT KILLS A $3,400/MONTH CLOUD STORAGE BILL the lenovo thinkstation pgx ships nvidia's gb10 grace blackwell superchip and 128gb of unified memory in a box the size of a mac mini at 1.2kg it runs an 80b qwen3 coder model at 25 to 40 tokens per second and a 196b step-3.5-flash moe model at 20 tokens per second locally the gb10 packs 6,144 cuda cores, 192 fifth-generation tensor cores and rates at 1 petaflop of fp4 with sparsity from a single 240 watt usb-c power supply fine tuning qwen 2.5 7b with lora took 18 minutes and 41gb of unified memory while the gpu pulled 65 watts and peaked at 77 degrees the box pulls a docker container from nvidia's registry and serves a frontier model on your local network with tool calling and zero data leaving your desk bookmark this and read the article below

AN AWS ENGINEER QUIETLY BUILT A 2 PETABYTE HOME SERVER FOR $9/MONTH THAT KILLS A $3,400/MONTH CLOUD STORAGE BILL the lenovo thinkstation pgx ships nvidia's gb10 grace blackwell superchip and 128gb of unified memory in a box the size of a mac mini at 1.2kg it runs an 80b qwen3 coder model at 25 to 40 tokens per second and a 196b step-3.5-flash moe model at 20 tokens per second locally the gb10 packs 6,144 cuda cores, 192 fifth-generation tensor cores and rates at 1 petaflop of fp4 with sparsity from a single 240 watt usb-c power supply fine tuning qwen 2.5 7b with lora took 18 minutes and 41gb of unified memory while the gpu pulled 65 watts and peaked at 77 degrees the box pulls a docker container from nvidia's registry and serves a frontier model on your local network with tool calling and zero data leaving your desk bookmark this and read the article below

starmex

193,839 Aufrufe • vor 4 Monaten

Do you want to own part of a AAA game? I know, you hear it all the time. “Triple A game”, you go to play it, it’s crap. This is different, and it’s only possible with Sonic (Sonic) speed, transaction cost, and of-course FeeM. A game that includes talent from Kojima, Ubisoft, EA Sports, Gameloft & more with advisors from NVIDIA. A game that you’ll be able to play on mobile, desktop, and then Xbox and PlayStation (yes really)! YES! A PRETTY BIG DEAL! Before I tell you about the sale, let me at least tell you about this game (being a massive gamer nerd, this excited me), so…. Introducing Animera (Search for Animera): • Fast-paced skill-based PvP in the Nubera galaxy • Compete in real-time space battles for real rewards It will be powered with $STRIKE: • Compete2Earn: win matches, earn tokens • Play2Burn: 5% of $STRIKE used in matches gets burned Oh, and with 8.75% of all game revenue will be used to buy & burn $SWPx, so the SwapX (SwapX) community owns a real stake in this AAA title. Absolutely insane. > Now let me tell you about its beta run quickly: • 16K+ beta signups • 500+ players added weekly • 7.5K+ matches already played • Launching to 500K+ mobile users via Nomina Games > How can you own a piece of Animera? June 5th at 2pm EDT the sale will go live on SwapX, it will go in three phases each lasting 12 hours or until sold out: PHASE 1️⃣: xNFT Holders Early access with exclusive perks and bonuses. These are for xNFT holders only you can get these here on paintswap PHASE 2️⃣ Whitelisted Communities These will be whitelisted from Creo Engine, SFA AGC, derp, and GOGLZ | SONIC 🥽💥. PHASE 3️⃣ Public Round Any remaining allocation will open to the public - only if Phases 1 & 2 don’t sell out. > What is the raise? Token Price & Allocation: • Token: $STRIKE • Currency: USDC • Total tokens for sale: 101.75M Unlock structure: • 50% unlocked at TGE • Remaining 50% claimable in 30 days • Raise cap: Max $100,000 per user, capped at $10,000 per xNFT • Purchase window priority: xNFT holders get early access (see above)! Transparency is key: Why I love working with the team is because transparency is crucial, so I’m going to tell you about its tokenomics, seed, and fully diluted valuation here: Token Symbol: STRIKE Total Supply: 370,000,000 Initial FDV: $1.48M Total Raise: $950,160 Total Initial Unlock: 112,947,501 STRIKE Initial Market Cap (excluding liquidity): $303,790 Token Allocation: • Seed Round: 59.2M tokens (16% allocation), with a 1-month cliff and linear vesting over 9 months. • Private Round: 94.35M tokens (25.5% allocation), with a 1-month cliff and 6-month vesting period. • Crowdsale: 10.75M tokens (2.91% allocation), unlocked 50% at TGE. • xNFT Holders: 10M tokens (2.7% allocation), with a 1-month cliff. • Liquidity: 37M tokens (10% allocation), with no lock or vesting. • Team: 18.5M tokens (5% allocation), with a 6-month cliff and 12-month vesting. • Rewards: 28.6M tokens (8% allocation), vested over 18 months. • Product Growth: 19.6M tokens (5.3% allocation), vested over 24 months. Token Offering: • Seed Round: Priced at $0.0033 per token, raising $195,360 by selling 59.2M tokens. 10% unlocks at TGE, with a 1-month cliff and 9-month vesting. The initial market cap from seed unlock is $234,127. • Private Round: Priced at $0.0037 per token, raising $349,095 for 94.35M tokens. 15% unlocks at TGE, with a 1-month cliff and 6-month vesting. Initial market cap contribution is $262,508. • Crowdsale: Priced at $0.0040 per token, raising $407,000 by selling 10.75M tokens. 50% unlocks at TGE, with no cliff or vesting. Adds $283,790 to the initial market cap. It’s important you had the full information at hand so you can decide whether or not you’d like to participate. I will be, because it’s a low FDV and it looks great. This is not financial advice, I’m helping the team out. Below is real gameplay: Further details: 👇

Do you want to own part of a AAA game? I know, you hear it all the time. “Triple A game”, you go to play it, it’s crap. This is different, and it’s only possible with Sonic (Sonic) speed, transaction cost, and of-course FeeM. A game that includes talent from Kojima, Ubisoft, EA Sports, Gameloft & more with advisors from NVIDIA. A game that you’ll be able to play on mobile, desktop, and then Xbox and PlayStation (yes really)! YES! A PRETTY BIG DEAL! Before I tell you about the sale, let me at least tell you about this game (being a massive gamer nerd, this excited me), so…. Introducing Animera (Search for Animera): • Fast-paced skill-based PvP in the Nubera galaxy • Compete in real-time space battles for real rewards It will be powered with $STRIKE: • Compete2Earn: win matches, earn tokens • Play2Burn: 5% of $STRIKE used in matches gets burned Oh, and with 8.75% of all game revenue will be used to buy & burn $SWPx, so the SwapX (SwapX) community owns a real stake in this AAA title. Absolutely insane. > Now let me tell you about its beta run quickly: • 16K+ beta signups • 500+ players added weekly • 7.5K+ matches already played • Launching to 500K+ mobile users via Nomina Games > How can you own a piece of Animera? June 5th at 2pm EDT the sale will go live on SwapX, it will go in three phases each lasting 12 hours or until sold out: PHASE 1️⃣: xNFT Holders Early access with exclusive perks and bonuses. These are for xNFT holders only you can get these here on paintswap PHASE 2️⃣ Whitelisted Communities These will be whitelisted from Creo Engine, SFA AGC, derp, and GOGLZ | SONIC 🥽💥. PHASE 3️⃣ Public Round Any remaining allocation will open to the public - only if Phases 1 & 2 don’t sell out. > What is the raise? Token Price & Allocation: • Token: $STRIKE • Currency: USDC • Total tokens for sale: 101.75M Unlock structure: • 50% unlocked at TGE • Remaining 50% claimable in 30 days • Raise cap: Max $100,000 per user, capped at $10,000 per xNFT • Purchase window priority: xNFT holders get early access (see above)! Transparency is key: Why I love working with the team is because transparency is crucial, so I’m going to tell you about its tokenomics, seed, and fully diluted valuation here: Token Symbol: STRIKE Total Supply: 370,000,000 Initial FDV: $1.48M Total Raise: $950,160 Total Initial Unlock: 112,947,501 STRIKE Initial Market Cap (excluding liquidity): $303,790 Token Allocation: • Seed Round: 59.2M tokens (16% allocation), with a 1-month cliff and linear vesting over 9 months. • Private Round: 94.35M tokens (25.5% allocation), with a 1-month cliff and 6-month vesting period. • Crowdsale: 10.75M tokens (2.91% allocation), unlocked 50% at TGE. • xNFT Holders: 10M tokens (2.7% allocation), with a 1-month cliff. • Liquidity: 37M tokens (10% allocation), with no lock or vesting. • Team: 18.5M tokens (5% allocation), with a 6-month cliff and 12-month vesting. • Rewards: 28.6M tokens (8% allocation), vested over 18 months. • Product Growth: 19.6M tokens (5.3% allocation), vested over 24 months. Token Offering: • Seed Round: Priced at $0.0033 per token, raising $195,360 by selling 59.2M tokens. 10% unlocks at TGE, with a 1-month cliff and 9-month vesting. The initial market cap from seed unlock is $234,127. • Private Round: Priced at $0.0037 per token, raising $349,095 for 94.35M tokens. 15% unlocks at TGE, with a 1-month cliff and 6-month vesting. Initial market cap contribution is $262,508. • Crowdsale: Priced at $0.0040 per token, raising $407,000 by selling 10.75M tokens. 50% unlocks at TGE, with no cliff or vesting. Adds $283,790 to the initial market cap. It’s important you had the full information at hand so you can decide whether or not you’d like to participate. I will be, because it’s a low FDV and it looks great. This is not financial advice, I’m helping the team out. Below is real gameplay: Further details: 👇

hoeem

21,634 Aufrufe • vor 1 Jahr

Holy shit... Microsoft open sourced an inference framework that runs a 100B parameter LLM on a single CPU. It's called BitNet. And it does what was supposed to be impossible. No GPU. No cloud. No $10K hardware setup. Just your laptop running a 100-billion parameter model at human reading speed. Here's how it works: Every other LLM stores weights in 32-bit or 16-bit floats. BitNet uses 1.58 bits. Weights are ternary just -1, 0, or +1. That's it. No floats. No expensive matrix math. Pure integer operations your CPU was already built for. The result: - 100B model runs on a single CPU at 5-7 tokens/second - 2.37x to 6.17x faster than llama.cpp on x86 - 82% lower energy consumption on x86 CPUs - 1.37x to 5.07x speedup on ARM (your MacBook) - Memory drops by 16-32x vs full-precision models The wildest part: Accuracy barely moves. BitNet b1.58 2B4T their flagship model was trained on 4 trillion tokens and benchmarks competitively against full-precision models of the same size. The quantization isn't destroying quality. It's just removing the bloat. What this actually means: - Run AI completely offline. Your data never leaves your machine - Deploy LLMs on phones, IoT devices, edge hardware - No more cloud API bills for inference - AI in regions with no reliable internet The model supports ARM and x86. Works on your MacBook, your Linux box, your Windows machine. 27.4K GitHub stars. 2.2K forks. Built by Microsoft Research. 100% Open Source. MIT License.

Holy shit... Microsoft open sourced an inference framework that runs a 100B parameter LLM on a single CPU. It's called BitNet. And it does what was supposed to be impossible. No GPU. No cloud. No $10K hardware setup. Just your laptop running a 100-billion parameter model at human reading speed. Here's how it works: Every other LLM stores weights in 32-bit or 16-bit floats. BitNet uses 1.58 bits. Weights are ternary just -1, 0, or +1. That's it. No floats. No expensive matrix math. Pure integer operations your CPU was already built for. The result: - 100B model runs on a single CPU at 5-7 tokens/second - 2.37x to 6.17x faster than llama.cpp on x86 - 82% lower energy consumption on x86 CPUs - 1.37x to 5.07x speedup on ARM (your MacBook) - Memory drops by 16-32x vs full-precision models The wildest part: Accuracy barely moves. BitNet b1.58 2B4T their flagship model was trained on 4 trillion tokens and benchmarks competitively against full-precision models of the same size. The quantization isn't destroying quality. It's just removing the bloat. What this actually means: - Run AI completely offline. Your data never leaves your machine - Deploy LLMs on phones, IoT devices, edge hardware - No more cloud API bills for inference - AI in regions with no reliable internet The model supports ARM and x86. Works on your MacBook, your Linux box, your Windows machine. 27.4K GitHub stars. 2.2K forks. Built by Microsoft Research. 100% Open Source. MIT License.

Guri Singh

2,180,357 Aufrufe • vor 7 Monaten

Today, we’re launching deel Founder Hours. 1-on-1 sessions with our exec team to help founders tackle the problems you can’t Google (ask Chatgpt?). As you scale, the hardest decisions don’t come with playbooks. They come from people who’ve been there. Deel is now 7,000+ people, 15+ products, and $1B+ in revenue - but none of that happened alone. We leaned heavily on founders, operators, and investors who gave us their time and hard-earned lessons. Now it’s time to give back! Twice a year, Deel execs (COO, CTO, CRO, and more) will spend a full day talking directly with founders ~10–20 deep, honest conversations. No pitches. Just real problems, real stories, and practical advice. I’m kicking it off and hosting the first Founder Hours on Jan 5. If you’re wrestling with something hard and want a 1-on-1, apply in the comments - We won't be able to help everyone but will do our best! Let’s pay it forward 💜

Today, we’re launching deel Founder Hours. 1-on-1 sessions with our exec team to help founders tackle the problems you can’t Google (ask Chatgpt?). As you scale, the hardest decisions don’t come with playbooks. They come from people who’ve been there. Deel is now 7,000+ people, 15+ products, and $1B+ in revenue - but none of that happened alone. We leaned heavily on founders, operators, and investors who gave us their time and hard-earned lessons. Now it’s time to give back! Twice a year, Deel execs (COO, CTO, CRO, and more) will spend a full day talking directly with founders ~10–20 deep, honest conversations. No pitches. Just real problems, real stories, and practical advice. I’m kicking it off and hosting the first Founder Hours on Jan 5. If you’re wrestling with something hard and want a 1-on-1, apply in the comments - We won't be able to help everyone but will do our best! Let’s pay it forward 💜

Alex Bouaziz

35,766 Aufrufe • vor 9 Monaten

Silencio has surpassed 1 million on-chain devices on peaq This is the live count of real contributors using their smartphones to capture and stream the physical world into digital infrastructure. Each device is opt-in, privacy-preserving, and individually rewarded in $SLC. Together, these contributors have recorded more than 10.6 million hours of real-world acoustic signal. The network spans 190 countries and has generated over 38.8 Billion environmental datapoints. All of it is live. All of it is on-chain. Silencio represents a deployed DePIN system operating at global scale. Its data is already integrated into major enterprise marketplaces. Its token economy is active, deflationary, and fully aligned with participation. This milestone is not just about numbers. It is about trust, coordination, and measurable impact. It proves that decentralized networks can scale without compromising on quality, compliance, or utility. Built on peaq. Secured by $SLC and the Blocksound Foundation. Powered by people. This is not the future of data infrastructure. It is already in motion.

Silencio has surpassed 1 million on-chain devices on peaq This is the live count of real contributors using their smartphones to capture and stream the physical world into digital infrastructure. Each device is opt-in, privacy-preserving, and individually rewarded in $SLC. Together, these contributors have recorded more than 10.6 million hours of real-world acoustic signal. The network spans 190 countries and has generated over 38.8 Billion environmental datapoints. All of it is live. All of it is on-chain. Silencio represents a deployed DePIN system operating at global scale. Its data is already integrated into major enterprise marketplaces. Its token economy is active, deflationary, and fully aligned with participation. This milestone is not just about numbers. It is about trust, coordination, and measurable impact. It proves that decentralized networks can scale without compromising on quality, compliance, or utility. Built on peaq. Secured by $SLC and the Blocksound Foundation. Powered by people. This is not the future of data infrastructure. It is already in motion.

Silencio | Voice Data for AI

22,863 Aufrufe • vor 1 Jahr

Jointly announcing EAGLE-3 with SGLang: Setting a new record in LLM inference acceleration! - 5x🚀than vanilla (on HF) - 1.4x🚀than EAGLE-2 (on HF) - A record of ~400 TPS on LLama 3.1 8B with a single H100 (on SGLang) - 1.65x🚀in latency even for large bs=64 (on SGLang) - A new scaling law: more training data, better speedup - Apache 2.0 Paper: Code: SGLang version: ⚒️Takeaway: Introducing training-time test, a novel draft model training technique: we replace feature prediction with direct token prediction and shift from top-layer-only features to multi-layer feature fusion. This approach unlocks a new scaling law previously undiscovered in EAGLE and EAGLE-2. 🙏Acknowledge: We would like to thank the SGLang team (zhyncs Lianmin Zheng Ying Sheng James Liu, Ke Bao, and others LMSYS Org) for their merge and careful evaluation of EAGLE-3 on SGLang. 🤝Want to collaborate? We're a small academic group with limited GPU resources. If you're interested in supporting our next version of EAGLE or would like us to train a preliminary version tailored to a specific model, please get in touch! Joint work with Yuhui Li, Fangyun Wei, and Chao Zhang

Jointly announcing EAGLE-3 with SGLang: Setting a new record in LLM inference acceleration! - 5x🚀than vanilla (on HF) - 1.4x🚀than EAGLE-2 (on HF) - A record of ~400 TPS on LLama 3.1 8B with a single H100 (on SGLang) - 1.65x🚀in latency even for large bs=64 (on SGLang) - A new scaling law: more training data, better speedup - Apache 2.0 Paper: Code: SGLang version: ⚒️Takeaway: Introducing training-time test, a novel draft model training technique: we replace feature prediction with direct token prediction and shift from top-layer-only features to multi-layer feature fusion. This approach unlocks a new scaling law previously undiscovered in EAGLE and EAGLE-2. 🙏Acknowledge: We would like to thank the SGLang team (zhyncs Lianmin Zheng Ying Sheng James Liu, Ke Bao, and others LMSYS Org) for their merge and careful evaluation of EAGLE-3 on SGLang. 🤝Want to collaborate? We're a small academic group with limited GPU resources. If you're interested in supporting our next version of EAGLE or would like us to train a preliminary version tailored to a specific model, please get in touch! Joint work with Yuhui Li, Fangyun Wei, and Chao Zhang

Hongyang Zhang

42,597 Aufrufe • vor 1 Jahr

Qwen3.8-Flash-Next now reaches ~43 tok/s after a 122,902-token prompt on ONE DGX Spark. ⚡🚀 MTP k=2 won my draft-depth sweep, with +42.5% mean decode over no draft. The PLE table stays fully on-device. I promised the deeper MTP tests. Here are the results, and now you can explore them in an interactive benchmark page too. 𝗧𝗪𝗢 𝗗𝗥𝗔𝗙𝗧 𝗧𝗢𝗞𝗘𝗡𝗦 𝗪𝗢𝗡 Mean single-request decode with 32K context configured: MTP k=2: 39.21 tok/s MTP k=3: 36.42 tok/s MTP k=1: 35.18 tok/s No draft: 27.51 tok/s k=2 also produced the fastest individual sweep run: 41.34 tok/s. Four runs each for no draft, k=1 and k=2. Seven for k=3. Decode excludes time to first token. Here, k means speculative draft depth, not quantization bits. k=3 produced more tokens per step, but the extra drafting work did not pay off in throughput. k=2 is my current pick for this setup. 𝗧𝗛𝗘 𝟭𝟮𝟯𝗞-𝗧𝗢𝗞𝗘𝗡 𝗣𝗥𝗢𝗠𝗣𝗧 𝗧𝗘𝗦𝗧 I then ran a separate long-prompt comparison: Actual input: 122,902 tokens Configured context: 262,144 Requested output: 128 tokens One request at a time MTP k=2: ~43 tok/s No draft: 26.2 tok/s Time to first token: 110.6 seconds with MTP 107.0 seconds without it The win here is generation speed, not faster prefill. To keep the scope clear: 256K was the configured limit. This was a real ~123K input, not a completely filled 256K window or a full k sweep at that depth. 𝗣𝗟𝗘 𝗦𝗧𝗔𝗬𝗦 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 Whole model on-device: 78.57 GiB Packed 5-bit PLE table: 30.4 GiB, included in that total No NVMe PLE offload in this build. This is still turboderp’s 3.05bpw_h5_ng5 EXL3 pack, served through my vllm-exl3 integration. My work here is the serving integration and testing. These are preliminary performance measurements, not a quality evaluation or a claim of bit-exact full-output parity. 𝗘𝗫𝗣𝗟𝗢𝗥𝗘 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 The benchmark page has the individual sweep values, long-prompt comparison, and measurement scope. You can play the animation, export the charts, or download the HTML and data to render them yourself. No Spark needed to view the results. Credit to turboderp / ExLlamaV3 for the pack and kernels, vLLM for the serving engine, and Qwen Qwen Developers for the model. Recipe + reproduction: Interactive benchmark:

Qwen3.8-Flash-Next now reaches ~43 tok/s after a 122,902-token prompt on ONE DGX Spark. ⚡🚀 MTP k=2 won my draft-depth sweep, with +42.5% mean decode over no draft. The PLE table stays fully on-device. I promised the deeper MTP tests. Here are the results, and now you can explore them in an interactive benchmark page too. 𝗧𝗪𝗢 𝗗𝗥𝗔𝗙𝗧 𝗧𝗢𝗞𝗘𝗡𝗦 𝗪𝗢𝗡 Mean single-request decode with 32K context configured: MTP k=2: 39.21 tok/s MTP k=3: 36.42 tok/s MTP k=1: 35.18 tok/s No draft: 27.51 tok/s k=2 also produced the fastest individual sweep run: 41.34 tok/s. Four runs each for no draft, k=1 and k=2. Seven for k=3. Decode excludes time to first token. Here, k means speculative draft depth, not quantization bits. k=3 produced more tokens per step, but the extra drafting work did not pay off in throughput. k=2 is my current pick for this setup. 𝗧𝗛𝗘 𝟭𝟮𝟯𝗞-𝗧𝗢𝗞𝗘𝗡 𝗣𝗥𝗢𝗠𝗣𝗧 𝗧𝗘𝗦𝗧 I then ran a separate long-prompt comparison: Actual input: 122,902 tokens Configured context: 262,144 Requested output: 128 tokens One request at a time MTP k=2: ~43 tok/s No draft: 26.2 tok/s Time to first token: 110.6 seconds with MTP 107.0 seconds without it The win here is generation speed, not faster prefill. To keep the scope clear: 256K was the configured limit. This was a real ~123K input, not a completely filled 256K window or a full k sweep at that depth. 𝗣𝗟𝗘 𝗦𝗧𝗔𝗬𝗦 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 Whole model on-device: 78.57 GiB Packed 5-bit PLE table: 30.4 GiB, included in that total No NVMe PLE offload in this build. This is still turboderp’s 3.05bpw_h5_ng5 EXL3 pack, served through my vllm-exl3 integration. My work here is the serving integration and testing. These are preliminary performance measurements, not a quality evaluation or a claim of bit-exact full-output parity. 𝗘𝗫𝗣𝗟𝗢𝗥𝗘 𝗧𝗛𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 The benchmark page has the individual sweep values, long-prompt comparison, and measurement scope. You can play the animation, export the charts, or download the HTML and data to render them yourself. No Spark needed to view the results. Credit to turboderp / ExLlamaV3 for the pack and kernels, vLLM for the serving engine, and Qwen Qwen Developers for the model. Recipe + reproduction: Interactive benchmark:

Cruz

12,258 Aufrufe • vor 1 Monat

how to grow your product from $0 to $10k MRR with AI influencers (using GPT Images 2.0 and Seedance 2.0): > start on Pinterest, save 15-20 reference bodies that fit your niche, after body type, lighting, styling, vibe, the raw material for an avatar you'll run for months > drop the refs into GPT Images 2.0 on thinking mode, have it extract the visual profile and generate a character sheet (front, side, 3-quarter, locked outfit, locked lighting), this sheet is the only asset that carries across every clip you ever ship > feed it into Higgsfield marketing studio as your avatar > paste your product link, prompt a 15-second UGC for tiktok, Seedance hands back a video with perfect lipsync and native audio in one pass > repeat that whole workflow per influencer, post daily per account, slip a product pitch in every 5-7 uploads, done and yes, you can literally create an army of influencers, but you'll need a proper system... lemme show you how it's done you need to factory reset an iphone, here's how to set it up: > set language and region to your target country (most likely US) > fresh gmail and fresh apple ID per device > match the app store region > run a paid dedicated-IP VPN > 2-3 accounts per device max, warm each one 1-2 days before the first post > VPN on at all times this is how you nail distribution at scale for cheap

Machina

21,283 Aufrufe • vor 5 Monaten

🚨 THE BIGGEST BOTTLENECK IN AI ISN'T COMPUTING POWER ANYMORE IT'S MOVING DATA. Instead of laying new cables, Chinese researchers have upgraded existing fiber infrastructure by doing two things at once: Using three wavelength bands (C + L + S) instead of the usual two. Using four cores inside each fiber instead of one. Each core acts like an independent highway, and each band acts like an extra lane on that highway. Together, they’ve reportedly increased transmission capacity per core by nearly 50% and overall data throughput by up to 5×. This matters enormously for AI. Modern AI clusters move terabits of data per second between thousands of GPUs. The biggest bottleneck is often not the chips themselves, but moving data fast enough between them. If you can push 5× more data through the same physical cables, you can train bigger models faster and reduce network congestion. Why this is significant: • It shows multi-core + extended spectrum technology moving from labs into real-world commercial use • The system has already run over 35 km of existing telecom network • It could be especially useful for submarine cables and large-scale data center interconnects • China is also eyeing it for its “Eastern Data, Western Computing” project The deeper implication: We’re reaching the physical limits of how much data we can push through single-core fibers using traditional methods. By combining spatial multiplexing (multiple cores) with spectral multiplexing (more wavelength bands), engineers are finding new ways to keep scaling bandwidth without having to dig up the planet to lay new cables. This kind of breakthrough is quiet but foundational it’s the kind of infrastructure upgrade that will determine how fast AI and cloud computing can actually grow in the coming years. The future of data movement might not require more cables. It might just require smarter ones. How important do you think multi-core and multi-band fiber will be for keeping up with AI’s exploding data demands? Follow for more frontier networking, photonics, and infrastructure technology.

🚨 THE BIGGEST BOTTLENECK IN AI ISN'T COMPUTING POWER ANYMORE IT'S MOVING DATA. Instead of laying new cables, Chinese researchers have upgraded existing fiber infrastructure by doing two things at once: Using three wavelength bands (C + L + S) instead of the usual two. Using four cores inside each fiber instead of one. Each core acts like an independent highway, and each band acts like an extra lane on that highway. Together, they’ve reportedly increased transmission capacity per core by nearly 50% and overall data throughput by up to 5×. This matters enormously for AI. Modern AI clusters move terabits of data per second between thousands of GPUs. The biggest bottleneck is often not the chips themselves, but moving data fast enough between them. If you can push 5× more data through the same physical cables, you can train bigger models faster and reduce network congestion. Why this is significant: • It shows multi-core + extended spectrum technology moving from labs into real-world commercial use • The system has already run over 35 km of existing telecom network • It could be especially useful for submarine cables and large-scale data center interconnects • China is also eyeing it for its “Eastern Data, Western Computing” project The deeper implication: We’re reaching the physical limits of how much data we can push through single-core fibers using traditional methods. By combining spatial multiplexing (multiple cores) with spectral multiplexing (more wavelength bands), engineers are finding new ways to keep scaling bandwidth without having to dig up the planet to lay new cables. This kind of breakthrough is quiet but foundational it’s the kind of infrastructure upgrade that will determine how fast AI and cloud computing can actually grow in the coming years. The future of data movement might not require more cables. It might just require smarter ones. How important do you think multi-core and multi-band fiber will be for keeping up with AI’s exploding data demands? Follow for more frontier networking, photonics, and infrastructure technology.

TheNewPhysics

20,485 Aufrufe • vor 4 Monaten

Probably the hackiest "bypasses" I've ever pulled off in botting: The Shopify REST API was for a long time the best and fastest way to checkout on Shopify. High rate limits, no captchas, no checkpoint and easy to work with. This is what gave Cyber the edge at the very start in 2018 (Dashe days) The Shopify team (with Jean-Michel Lemieux at the helm) spent years plugging the holes to make sure it wasn't used during drops. Around 2020, Shopify released Shopify Pay, which allowed you to expedite checkouts on all stores by linking your details to a phone number. Turned out the payment tokens generated through there allowed you to use the payments API as you could 2 years back. We generated around 3k Shop Pay accounts with randomised details, load balanced them across Cyber tasks (as rate limits were per account) and dominated every single release (...until it got patched). We called it Fast2. 3 years later I noticed that the payments API could still be used with Shopify mobile apps using Apple Pay. Turns out you could still send the users' tokenised credit card data in the request and that would take precedence - and as long as a valid signed Apple Pay token generated in the last 2 minutes was also in the request data, the payment request would go through and bypass captcha and checkpoint. Generating those Apple Pay tokens was the challenge, and reversing iOS frameworks was not an option. So I repurposed an old iPhone with a custom sideloaded Apple Pay app that would submit the token to our servers when the flow completed. And by combining accessibility macros, developer sandbox mode to prevent using real credit cards, a voice activation trigger and a old pair of earphones taped onto the microphone - a makeshift Apple Pay token server was born and Fast2 (or was it Fast3?) was back. Jean-Michel Lemieux thank you for providing me with the longest, most multi faceted puzzle ever.

Probably the hackiest "bypasses" I've ever pulled off in botting: The Shopify REST API was for a long time the best and fastest way to checkout on Shopify. High rate limits, no captchas, no checkpoint and easy to work with. This is what gave Cyber the edge at the very start in 2018 (Dashe days) The Shopify team (with Jean-Michel Lemieux at the helm) spent years plugging the holes to make sure it wasn't used during drops. Around 2020, Shopify released Shopify Pay, which allowed you to expedite checkouts on all stores by linking your details to a phone number. Turned out the payment tokens generated through there allowed you to use the payments API as you could 2 years back. We generated around 3k Shop Pay accounts with randomised details, load balanced them across Cyber tasks (as rate limits were per account) and dominated every single release (...until it got patched). We called it Fast2. 3 years later I noticed that the payments API could still be used with Shopify mobile apps using Apple Pay. Turns out you could still send the users' tokenised credit card data in the request and that would take precedence - and as long as a valid signed Apple Pay token generated in the last 2 minutes was also in the request data, the payment request would go through and bypass captcha and checkpoint. Generating those Apple Pay tokens was the challenge, and reversing iOS frameworks was not an option. So I repurposed an old iPhone with a custom sideloaded Apple Pay app that would submit the token to our servers when the flow completed. And by combining accessibility macros, developer sandbox mode to prevent using real credit cards, a voice activation trigger and a old pair of earphones taped onto the microphone - a makeshift Apple Pay token server was born and Fast2 (or was it Fast3?) was back. Jean-Michel Lemieux thank you for providing me with the longest, most multi faceted puzzle ever.

Lucas

248,136 Aufrufe • vor 1 Monat

To replace animal testing with AI, we need MASSIVE human datasets. Today, we're thrilled to share Axiom's new data exploration tool, providing the ability to visually explore the world's largest primary human liver toxicity dataset. Built with Axiom's proprietary wetlab protocols, our dataset includes detailed liver toxicity profiles for over 100,000 distinct molecules. The key to this dataset is our ability to do high-throughput, multiplexed high-content screening with primary human liver cells. Traditionally, toxicity assays either sacrifice throughput or sacrifice biological relevance (using easy-to-grow immortalized cell lines instead of real human cells). We managed to combine throughput, physiological relevance, and multiplexing in one platform. The assays run in a high throughput format using automation, meaning thousands of compound-dose conditions can be tested in one experiment. We achieved this using pooled primary human hepatocytes, which are often fragile and expensive. By systemizing our automation and quality control processes, we were able to run over 120+ batches on the same donor pool with incredible reproducibility and consistency. We did this while integrating many readouts per well, whereas many existing toxicity assays only do a single readout. Our multiplexed approach provides far more data per experiment enabling us to measure 10-20 different toxicity phenotypes such as apoptosis, necrosis, mitochondrial fission, endoplasmic reticulum stress, stress granule formation, microtubules, and more all from a single well on a 384-well plate! The combination of scale, high content information, and data quality is exactly what is needed to train highly accurate AI models in biology. If you're interested, please explore the dataset in the comments below and let me know if you want to chat about the details!

To replace animal testing with AI, we need MASSIVE human datasets. Today, we're thrilled to share Axiom's new data exploration tool, providing the ability to visually explore the world's largest primary human liver toxicity dataset. Built with Axiom's proprietary wetlab protocols, our dataset includes detailed liver toxicity profiles for over 100,000 distinct molecules. The key to this dataset is our ability to do high-throughput, multiplexed high-content screening with primary human liver cells. Traditionally, toxicity assays either sacrifice throughput or sacrifice biological relevance (using easy-to-grow immortalized cell lines instead of real human cells). We managed to combine throughput, physiological relevance, and multiplexing in one platform. The assays run in a high throughput format using automation, meaning thousands of compound-dose conditions can be tested in one experiment. We achieved this using pooled primary human hepatocytes, which are often fragile and expensive. By systemizing our automation and quality control processes, we were able to run over 120+ batches on the same donor pool with incredible reproducibility and consistency. We did this while integrating many readouts per well, whereas many existing toxicity assays only do a single readout. Our multiplexed approach provides far more data per experiment enabling us to measure 10-20 different toxicity phenotypes such as apoptosis, necrosis, mitochondrial fission, endoplasmic reticulum stress, stress granule formation, microtubules, and more all from a single well on a 384-well plate! The combination of scale, high content information, and data quality is exactly what is needed to train highly accurate AI models in biology. If you're interested, please explore the dataset in the comments below and let me know if you want to chat about the details!

Brandon White

25,117 Aufrufe • vor 1 Jahr

Depth Any Video with Scalable Synthetic Data AI physicists and chemists continue to make strides in depth estimation from video. Check out this new paper featuring some impressive examples. See the thread for more details (unfortunately no code yet). Abstract: Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates 0 - even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency.

Depth Any Video with Scalable Synthetic Data AI physicists and chemists continue to make strides in depth estimation from video. Check out this new paper featuring some impressive examples. See the thread for more details (unfortunately no code yet). Abstract: Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse game environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates 0 - even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency.

MrNeRF

27,428 Aufrufe • vor 2 Jahren

it took us 3 weeks to get 500k+ views with a fully AI UGC on instagram and it taught me something huge if you create good content for long enough the algorithm is very simple it works to keep finding the type of user that engaged with your post similar to paid ads but this only works if you stay within niche and post consistently great content so stop seeing 1,000-4000 views as bad it’s simply taking the time to learn we’re being extremely aggressive with our marketing and created 40+ content variations each A/B testing a different part of the video - hook, music, transitions, wallpapers, workout, location, CTA, captions every single component is being tested we had a huge set back with Meta randomly banning 8 of our accounts last week we got the content ready and aim to post 20x per day minimum by 6/10 each day breaking down the data and creating more of what works then running it on paid ads both tiktok spark ads and meta ads marketing is definitely takes more time but at least shipping took just 2 weeks using Rork Max + Opus 4.7 at the time i launched, there was no other possible way to natively code in SwiftUI

it took us 3 weeks to get 500k+ views with a fully AI UGC on instagram and it taught me something huge if you create good content for long enough the algorithm is very simple it works to keep finding the type of user that engaged with your post similar to paid ads but this only works if you stay within niche and post consistently great content so stop seeing 1,000-4000 views as bad it’s simply taking the time to learn we’re being extremely aggressive with our marketing and created 40+ content variations each A/B testing a different part of the video - hook, music, transitions, wallpapers, workout, location, CTA, captions every single component is being tested we had a huge set back with Meta randomly banning 8 of our accounts last week we got the content ready and aim to post 20x per day minimum by 6/10 each day breaking down the data and creating more of what works then running it on paid ads both tiktok spark ads and meta ads marketing is definitely takes more time but at least shipping took just 2 weeks using Rork Max + Opus 4.7 at the time i launched, there was no other possible way to natively code in SwiftUI

Jah Mills

92,243 Aufrufe • vor 4 Monaten

$🚀 A better, faster co-folding-based binding affinity model. Predicting how tightly a drug candidate binds to its target is critical in drug discovery. It also requires massive computational resources. State-of-the-art models can take 20 seconds to a minute per prediction, impractical for the demands of large scale early-stage programs . 💠 Today, Recursion’s Valence Labs is releasing Nesso-1: the fastest open-source co-folding-based binding affinity model available. At 1 second per prediction, it’s roughly 20x faster than our previous collaboration on Boltz-2 while matching or surpassing its accuracy across public and internal benchmarks. By leveraging NVIDIA Healthcare cuEquivariance, we’ve been able to further accelerate both training and inference by an additional 2-3x. We look forward to continuing to improve Nesso-1 in collaboration with NVIDIA. Weights and code are fully open-sourced. The core architectural ideas behind Nesso-1 build on the insight that coarse-grained co-folding representations can match full-atom models for affinity prediction at a fraction of the cost. Nesso-1 is the first open implementation of this approach with no proprietary dependencies, trained entirely on public data, built to be reproducible and extensible. We’re already using Nesso-1 internally in active drug discovery programs. Fast, reliable affinity prediction at scale is foundational to the kind of autonomous design loops that define our vision for Autonomous Precision Design and Nesso-1 is a meaningful step toward that. 👉 Report: 👉 Github: 👉 HF:$

🚀 A better, faster co-folding-based binding affinity model. Predicting how tightly a drug candidate binds to its target is critical in drug discovery. It also requires massive computational resources. State-of-the-art models can take 20 seconds to a minute per prediction, impractical for the demands of large scale early-stage programs . 💠 Today, Recursion’s Valence Labs is releasing Nesso-1: the fastest open-source co-folding-based binding affinity model available. At 1 second per prediction, it’s roughly 20x faster than our previous collaboration on Boltz-2 while matching or surpassing its accuracy across public and internal benchmarks. By leveraging NVIDIA Healthcare cuEquivariance, we’ve been able to further accelerate both training and inference by an additional 2-3x. We look forward to continuing to improve Nesso-1 in collaboration with NVIDIA. Weights and code are fully open-sourced. The core architectural ideas behind Nesso-1 build on the insight that coarse-grained co-folding representations can match full-atom models for affinity prediction at a fraction of the cost. Nesso-1 is the first open implementation of this approach with no proprietary dependencies, trained entirely on public data, built to be reproducible and extensible. We’re already using Nesso-1 internally in active drug discovery programs. Fast, reliable affinity prediction at scale is foundational to the kind of autonomous design loops that define our vision for Autonomous Precision Design and Nesso-1 is a meaningful step toward that. 👉 Report: 👉 Github: 👉 HF:

Recursion

156,802 Aufrufe • vor 2 Monaten

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?

Alok

41,735 Aufrufe • vor 1 Monat