Qwen 3.6 is frontier for local. It also thinks... forever. I tried a dumb inference-time trick: make its block obey a tiny grammar. Result: - HumanEval+: 22x fewer think tokens, no accuracy loss - LiveCodeBench public slice: +14% pass@1, ~5x fewer total tokensshow more

/1 Gemma 4 31B just crushed Qwen 3.6 27B... in a local LLM gamedev contest inside atomic.chat (prompt is below) Device: MacBook Pro M5 Max, 64GB RAM Results: Qwen 3.6 27B: 32 tokens/sec · 18m 04s · 33,946 tokens Gemma 4 31B: 27 tokens/sec · 3m 51s · 6,209 tokens So what is more important: tokens per second, or the quality of the final answer? Qwen made a very long response and showed more creativity and visual style. But Gemma gave a shorter, clearer, and more logical answer in much less time. In this one-shot Pac-Man gamedev contest, Gemma 4 31B was the clear winner. Its game logic was stronger: click reactions were smoother, and it handled interactions with elements like walls, ghosts, and particle effects better. But this was only one test. Maybe Qwen 3.6 27B can show better results with better settings. Open the comments, try our prompt, and share your result below.show more

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: 👇show more

Most recent diffusion language model research (that I’ve seen)... seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to *potentially* giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.show more

Qwen 3.8 27B Q4_K_M - 90 tokens/sec on a... single NVIDIA RTX 4090 (24 GB VRAM) with Dflash2! (MTP 60 tps -> 90 tps Dflash2!!!!) Local AI moves so fast (literally!) it’s terrifying. Z lab just dropped DFlash 2 for Qwen 3.8 27b and Muse Glimmer. I patched llama.cpp (PR #27342) and paired it with Unsloth’s Qwen 3.8 27B UD-Q4_K_XL quant. The result? Lossless 90 tokens/s decode. My last post highlighted native MTP hitting 60 t/s at 130,000 context. But DFlash 2 just completely shattered that ceiling. By using parallel block diffusion drafting (predicting whole blocks of tokens in a single pass using dynamic convolutions), DFlash achieves a massive 5.39 token acceptance rate. THE ALPHA TWEAK: `n-max 7` eats too much VRAM for draft states. But if you drop the draft limit to `--spec-draft-n-max 4`, you slash the VRAM overhead and actually increase the throughput. Here is the new 24GB VRAM Physics Matrix (DFlash 2 @ n-max 4): - 30k Context: 1,725 t/s prefill | 87.05 t/s decode | 22.2 GB VRAM - 80k Context: 1,789 t/s prefill | 84.20 t/s decode | 23.3 GB VRAM - 110k Context: 1,767 t/s prefill | 83.35 t/s decode | 23.96 GB VRAM (110k context at 83+ tokens a second sitting exactly on the 24GB hardware limit is absolute wizardry). How to compile the PR today: git clone cd llama.cpp git fetch origin pull/27342/head:pr-27342 git switch pr-27342 cmake -B build -DGGML_CUDA=ON && cmake --build build -j Llama.cpp flags for Dflash (110k Context Ceiling): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q4_K_M.gguf --spec-type draft-dflash --spec-draft-n-max 4 -c 110000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 The fact that the open source community is shipping block diffusion drafters so quickly that run entirely locally on a gaming GPU is unbelievable. If you own a single RTX 3090 or 4090, it is officially time to upgrade to qwen 3.8 27b with dflash 2 and cancel your API subscriptions and let local silicon eat the cloud. This model beats GPT 5.6 Terra, GLM 5.2 DeepSeek V4 Pro, Muse Spark 1.2 and Claude Opus 4.8 on the artificial analysis agentic index (details in the replies) Hugging Face GGUF links (Base + DFlash2) and the full visual VRAM scaling and Dflash2 vs MTP graphs are also in the replies below. are you sticking to native MTP for the 130k context, or sacrificing 20k context to redline your decode speed? How many tokens/sec are you pushing on your current local rig?show more

I told you to claim your free 16GB NVIDIA... GPU for learning Local LLMs. Now I’m going to show you how to double its inference speed without touching the hardware. Google Colab gives you an enterprise grade NVIDIA Tesla T4 GPU for free, roughly 4 hours every single day. It is the absolute perfect sandbox for learning AI engineering, testing inference flags, and pushing massive context windows. The local AI timeline is moving way too fast. If you aren't using Multi Token Prediction (MTP) yet, you are leaving massive performance on the table. I just pushed DeepMind’s Gemma 4 26B to 64.9 t/s on this exact free tier. Let's look at the raw benchmark data running on an Ubuntu Linux environment with the latest compiled llama.cpp binaries and quantized GGUFs from Unsloth via HuggingFace: # Qwen 3.5 9B (Dense): Base: [ Prompt: 626.7 t/s | Generation: 21.0 t/s ] With MTP: [ Prompt: 539.1 t/s | Generation: 24.8 t/s ] # Gemma 4 26B QAT (MoE): Base: [ Prompt: 634.2 t/s | Generation: 48.3 t/s ] With MTP: [ Prompt: 572.1 t/s | Generation: 64.9 t/s ] If you are paying attention, this single Colab notebook reveals 3 massive observations about the current state of local LLMs: # 1. The MTP Speedup (Software Overclocking) Standard autoregressive decoding guesses one token at a time. MTP acts like a highly optimized, built in speculative decoder. It predicts multiple future tokens at once and the main model verifies them in parallel. The result? Zero accuracy loss and a massive throughput increase. Gemma jumped from 48 to 65 t/s just by flipping a flag. # 2. The MoE Paradox (Bigger is Faster) How does a 26B parameter model absolutely destroy a 9B model in raw speed on the exact same hardware? Architecture. Qwen 3.5 9B is a dense model. it activates all 9 billion parameters for every single token. Gemma 4 26B is a Mixture of Experts (MoE) model. It routes data efficiently, activating only 4B parameters per token. You get the reasoning capabilities of a 26B model with the compute cost of a 4B model. 3. Thinking Efficiency When I ran the exact same complex prompt on both models, the larger MoE spent significantly fewer "thinking" tokens to arrive at the correct answer. A smarter model doesn't just give better answers; it gets to the point faster, saving you compute cycles and preserving your context window. # Want to run this yourself? Here are the exact llama.cpp CLI commands. For Qwen (MTP is baked into the main model): ./llama-cli -m Qwen3.5-9B-UD-Q4_K_XL.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 For Gemma (Using a separate lightweight draft model): ./llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --model-draft mtp-gemma-4-26B-A4B-it.gguf -p "Explain quantum computing." -n 2000 -c 8000 -ngl 99 -fa on --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Stop waiting for a $3,000 rig. Boot up Colab, pull these models, and start building your stack. I’ve put together a completely free, cell by cell Google Colab notebook that automates this entire workflow so you can test it yourself in 5 minutes and learn. Link to the notebook is in the comments below. Experiemt with different MTP parameters, context windows and post your results in the comments.show more

Day 11/90 of Inference Engineering How does vLLM work... and how is it used in production? Before we discuss how vLLM works internally, it helps to understand what vLLM is. At a high level, vLLM is an inference engine that is designed to serve LLMs to thousands of concurrent users efficiently while managing scarce compute and memory. The goal for vLLM is to maximize throughput and minimize latency; optimizing for the best inference economics and experience for end users. With every request from the end user, it eventually ends up in the engine core, gets scheduled alongside other requests from other concurrent users, executes on the GPU, and updates the KV cache with the new key and value vectors, and streams the tokens back to the user. The Scheduler decides what requests should execute next while continuously batching requests together to maximize GPU utilization. Continuous batching is an inference optimization that allows new requests to join a running batch as other requests finish generating tokens. This helps with keeping the GPU utilization high instead of letting it sit idle waiting for an entire batch to complete generating. After the scheduler dispatches the selected batch to the Model Executor, the Model Executor prepares the tensors and metadata required for inference, retrieves each request’s block table from KV Cache Manager, launches the optimized transformer forward pass on the GPU, computes the logits, updates the KV cache with the new key and value vectors, and finally returns the results for sampling and streaming. The KV Cache Manager uses the PagedAttention memory layout to allocate fixed-size cache blocks on demand and maintains a Free Block Queue on the CPU that tracks which blocks in the GPU’s Paged KV Cache are currently free. When a request needs additional KV cache space, the KV Cache manager takes a free block from the queue and assigns it to that request, thus avoiding an expensive search through GPU memory for available cache blocks. All of these components form the core of vLLM’s inference engine. The Scheduler determines what requests are executed, the Model Executor determines how those requests are executed, the KV Cache Manager determines where each request’s KV cache lives using the PagedAttention Memory Layout. This architecture enables vLLM to serve thousands of concurrent requests with high throughput, low latency, and efficient GPU memory utilization. Heres a little animation that visualizes everything! - I've also completed the forward pass for my mnist.c project. I had a nice chat with shrey birmiwal, such a knowledgeable guy. Excited to learn more about vLLM and implement a tiny-vLLM one day.show more

🧑‍🚀 Day 5 of the Cursor #vibejam Proudly sponsored... by Cursor + bolt.new + GLIF Today a little bit about Cursor 3, it's a completely rebuilt interface designed around agents, you can run multiple AI agents in parallel, hand off between local and cloud, and go from commit to merged PR. It's powered by Composer 2, their own frontier coding model that tops benchmarks at $0.50/M input tokens — and the Fast variant runs at 200 tokens/sec, more than 2x faster than other frontier models. Perfect for vibe coding your game jam entry. And edwin from Cursor is here to help you if you get stuck! Lots of activity again today again on here, the best games (or well previews) I saw today: 🍣 Sushi Belt by steeno 🌊 Surfing Simulator by Erik (this one is a basic demo but very promising as it has realistic water physics) 🦆 Grand Quack Auto by kyzo (or as I like to call it Bali Simulator) 🏛️ Roman Combat Sim by terry trusner (continued from yesterday because it's had a lot of progress I think) YOU HAVE 24 DAYS LEFT! Reply in this thread with updates on your current games to share your progress, and add tag #vibejam so I see and can include you in the daily tweet There's $35,000 in prizes for you to win, see threads below for more info. The Gold prize is $20,000, bronze is $10,000 and silver is $5,000! Wanna to participate? You can still start today and submit your game any time before May 1!show more

Introducing fx, a tiny, open, native coding agent from... Vercel Labs. Originally an internal tool, fx is a harness and CLI written in Zig, optimized for research and embedding in larger systems. Today, we're open sourcing it. fx is built on three principles: 1. Fast. A single native binary, no runtime to install. It cold starts in 10µs and does no unnecessary work or I/O before accepting input. fx is the answer to "how fast can a coding agent be?" 2. Light. The 6.3MiB binary uses single-digit megabytes of memory at baseline, made for instant installation and embedding in resource-constrained environments and agent sandboxes. 3. Open. Apache-2.0, model and provider agnostic, suitable for local and cloud inference. Its small core extends through skills, plugins, and MCP. Minimalism is an obsession throughout the entire harness: system prompt, tools, features, binary. The goal was to keep context usage and time to first token low, and make fx optimal for model benchmarking, sandboxing, evals, and gyms. You can use fx directly or embed it as infrastructure. The CLI feels more like a Unix shell than an IDE in the terminal: it preserves scroll history, produces minimal output, and uses complex TUI rendering very, very sparingly. Programmatically, 𝚏𝚡 𝚊𝚜𝚔 --𝚓𝚜𝚘𝚗 gives structured output, 𝚏𝚡 𝚊𝚌𝚙 connects to editors and other clients, and WebAssembly can even run the whole thing inside the browser (see: Privacy is a design constraint: no product telemetry, sessions and usage stay local, and no source code or prompts are shared with any endpoint other than inference. With local inference and auto-updates off, fx is fully hermetic. fx is experimental. Use at your own risk and expect frequent changes. Chat with us on X ( or file issues ( 𝚌𝚞𝚛𝚕 -𝚏𝚜𝚂𝙻 𝚏𝚡.𝚜𝚑/𝚜𝚎𝚝𝚞𝚙.𝚜𝚑 | 𝚋𝚊𝚜𝚑show more

1/7 Built a Polymarket trading bot over 3 months.... Here are the biggest mistakes that cost me real money > Went from v1 to v61. Every version fixed something painful. --- 2/7 Stop Loss killed more money than it saved. > Binary markets need room to breathe - fluctuations are normal. > Stop Loss was cutting positions on random noise and locking in losses right before the market flipped. > Removed it in v61. Immediately better. --- 3/7 Martingale + Stop Loss = a loss cascade. > Seemed logical: lost $5 -> bet $8, lost again -> bet $10. > In practice: a losing streak plus early exits = a hole in your balance in a single day. > Killed it. For good. --- 4/7 Smart Exit without Force Exit is a trap. > Token hits 90c (+75% profit), but the bot was waiting for a "BTC reversal" signal. > Market closes, token drops, profit gone. > Fix: hard Force Exit at 85c. No conditions, no waiting. --- 5/7 Blocking the 5:30-10:30 PM ET window felt safe. It wasn't. > NYSE open = sharp spikes = bad signals. Made sense to block it. > But the full block was also killing clean entries at 8-10:30 PM. > Had to split the zone into segments with different edge/move thresholds. --- 6/7 The Gamma API lies about market start time. > Start price ("price to beat") is the core input for every signal. > Gamma was returning stale data. Had to pull prices directly from Chainlink on-chain on Polygon. > That's its own adventure - polling a smart contract every 2 seconds at 2 AM. --- 7/7 The real lesson: don't overcomplicate what works. > v1: complex system, 10 indicators -> -$200/day > v61: "buy the expensive token for $5, exit at +30%" -> consistently green > Simpler logic = fewer failure points. > The bot runs 96 intervals a day. Every mistake shows up fast.show more

this is the worst local ai will ever be.... it only gets better from here. if you are not expanding your mind with these small models you are missing what's happening right now 99 percent tool call success rate. when steered well with the right skills and a framework like hermes agent the node becomes a cognition layer. not a chatbot. not a toy. an extension of how you think. i was cranking this node at 35 to 50 tok/s all day on personal experiments and now after all the work is done qwen 3.5 9B is iterating on its own code. the game it created. fixing its own bugs autonomously. and the part you should probably not miss is that all of this is happening on a RTX 3060. not an H100. not an A100. the card most of you have sitting in a drawer right now. if you just open that drawer and put that intelligence to work every tensor core on that card should be running for you. your work. your experiments. your thinking. you all have it but because nobody told you what this hardware can actually do in 2026 you never tried. the day it unlocks is the day you test your workload, understand the tradeoffs, debug the loops, and then decide if you need to scale the hardware. there is no point buying 3 mac studios when things done well you can squeeze a similar level of intelligence from 9B compared to 70B. but only when you create the right environment for your model through the right harness. and let me tell you i have tried claude code as a local harness. i have tried opencode. i have tried various others. somehow i landed on hermes agent and never left. there is something magical going on at Nous Research. the tool call parsers, the skills system, the way it handles small models natively. nothing else comes close for local inference. own your cognition. your AI. your agent. your prompts. your experiments. why give them away for free. those are who you are and they don't belong on someone else's servers being monitored. just give it a shot with your existing hardware. you run into a problem the community will help you. and if you are migrating from openclaw to hermes i will personally help you make the switch.show more

My dual RTX PRO 6000 setup is currently training... a Draft model for Qwen 3.6 27B! 🔥 I'm taking the paper DeepSeek dropped on 6/26 and going for a super ambitious application to the 27B scale. Thanks to my homelab, I was able to dive straight in — I read the paper and immediately started experimenting. The amount I've learned has been insane: - How memory bandwidth bottlenecks speed and clever ways to hack around it - Methods to train the draft model and boost its accuracy - Mechanisms to reference tokens all the way back to the previous one to skyrocket draft acceptance rates - The impact of Attention vs. GateDeltaNet on speculative decoding performance and how to handle those differences - The unique approaches and trade-offs of MTP, Dflash, JetSpec, and DSpark I could go on forever, but just from speculative decoding alone I've learned so much. The 27B architecture feels way more DSpark-native than JetSpec, so once draft training finishes, I'm going all-in with DSpark! My goal is to beat existing speculative decoding speeds outright — no task-specific shortcuts or cheating, pure general improvement. If you're into this kind of research, I'd love to hear your thoughts, impressions, and any suggestions — please reply! 🚀show more

I just ran Gemma 4 31B on @CerebrasSystems at... 1,800+ tokens/sec and it's multimodal. For context: that's 35x faster than a typical GPU endpoint, and the first token (reasoning included) lands in 1.5 seconds. This isn't a benchmark slide, I recorded the inference live. Prompt I used: "Create a simulation of an iPhone. Include at least one working dummy note taking app, a functional notification pulldown, high quality graphics, single HTML file, any libs via CDN." - Generation time: 3 seconds. - Notes app worked. - Notification panel worked. - Rendered first try. This is what wafer-scale inference unlocks, not just "faster," but a different category of product. When generation is this fast, you stop waiting and start iterating in real time. Why this matters: Gemma 4 31B is Google DeepMind's flagship open weight model, Apache 2.0 licensed, dense (not MoE), and built for efficiency over raw parameter count. It scores close to Claude Haiku 4.5 on the Artificial Analysis Intelligence Index (30 vs 29) but runs ~18x faster on Cerebras. It's also the first multimodal model on Cerebras's platform, meaning you can now feed it screenshots, documents, charts, and UI states at wafer scale speed. # Applications I'm most excited about: - Screenshot → Insight: Drop in a dashboard or document screenshot, get structured findings back instantly. no waiting, no batching. - Live UI generation: Full interactive interfaces (like my iPhone sim) generated and rendered in under 2 seconds. - Screenshot -> Patch: Feed it a broken UI + console error, get a minimal code fix and verification steps back. - Computer use & agentic loops: See -> reason -> act - verify, fast enough to keep a human in the loop instead of waiting on the model. - Long context summarization: Full research reports condensed into decision ready summaries you can read and requery in one sitting. The bigger unlock isn't the speed number itself, it's that agentic and multimodal loops (see -> reason -> output -> tool call -> verify -> retry) finally run in real time instead of feeling sluggish. As Logan Kilpatrick (Logan Kilpatrick) put it: "If every model was doing 2,000 tokens per second, you wouldn't build the same product and just have it be faster, you'd build different products." Gemma 4 31B is live now on Cerebras Inference Cloud in public preview. If you're building multimodal, agentic, or real time apps, this is worth testing today. What would you build with such insane inference throughput?show more

I just got Gemma 4 26B A4B MoE model... running fully locally with Hermes agent on an 8GB RTX 4060 and it's now backtesting trading strategies end to end, no hand holding. If you’re a trader or work on Wall Street, you don’t want to miss this. Yes. fully automated. No cloud. No APIs beyond market data. # Here's what I did: Setup: - Model: Gemma 4 26B-A4B QAT (MoE), Q4_K_XL Unsloth's quant (link in the comments) - Inference: llama.cpp (turboquant fork by Tom Turney link in the comments) - Hardware: RTX 4060, 8GB VRAM + 16GB RAM only (with 50 other chrome tabs open) - Context: 64K llama.cpp turboquant flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 64000 --cache-type-k q8_0 --cache-type-v turbo3 --port 8080 turboquant helps achieve high prefill and decode throughput for interactive sessions. throughput with Hermes agent: decode: 25+ tokens/sec prefill: 250+ tokens/sec # Then I gave the agent one task: Backtest a strategy: - Buy when RSI crosses above 30 - Sell at +2% profit or -1% stoploss - No overlapping positions - Use Google stock via yfinance - Generate a full HTML report with candlestick charts + signals What happened next was wild. It didn't just write code, it ran the entire workflow itself: Audited the environment (pip list, dependency check) Hit a ModuleNotFoundError, multiple Python installs were conflicting Ran where python to map every interpreter on the system Manually selected the correct Python 3.13 path and re ran the script Wrote a clean statevmachine backtester (strict no overlapping trades logic) Patched a yfinance MultiIndex quirk that would've crashed the script Built Plotly candlestick + RSI charts with buy/sell markers Calculated win rate, PnL, and summary stats Exported a polished single file HTML report. check the report at the end of the video or in the comments. Biggest takeaway: local LLMs aren't just "chat assistants" anymore. They debug their own environment, write production code, and ship a finished deliverable on consumer hardware, for $0 in API costs. If you're still calling local models "toys," you're already behind. This is just the beginning. Hermes agent just surpassed 1 trillion tokens in a single day on OpenRouter. Think about the scale of total token generation happening right now. Disclaimer: This is not financial advice. Consult a professional before making any trading decisions.show more

I am the Under Secretary for Strategic Resource Integration.... We sit inside the Executive Office of the President. You won't find us on an org chart. We requested it that way. You think an office like mine spins up in a crisis. We've been open since January. Venezuela wasn't a war. Venezuela was the beta. Maduro proved the model the morning we took delivery of him. The President called it "working out brilliantly for both." He's right. It works brilliantly for both of us, and the customer will come around to the "both" on their own time. Kharg Island is the rollout. A third the size of Manhattan, and 90 percent of Iran's crude runs through it. 2 million barrels a day, one valve. I have prayed my whole career for a country that legible. 53 billion in revenue. 11 percent of their GDP. All of it funneling through a rock you could walk across before lunch. I kept the Venezuela playbook. I changed the classification banner and let find-and-replace do Caracas to Tehran. That was the plan. That was the whole plan. Page 14 still says Caracas. No one on the distribution list has ever reached page 14. They initial page 1. Page 1 is projected revenue. It goes up and to the right, the only direction this town can read. The Energy Secretary testified he was "unaware." Under oath, to Congress, the man who runs American energy. Good. He was never read into the compartment. The fewer names on the account, the cleaner the account. He runs the department. He does not run this. We don't say invasion. We say onboarding. Every onboarding has friction. The Marines off the coast are not an occupying force. They're field engagement. Iran said an invasion of its islands would "shatter all restraint." I logged it as customer feedback and filed it. The file is full. Engaged customers submit the most feedback. It's the quiet ones who churn. And Iran can't churn. That's the part I'm proudest of. We built the account with no exit. No competitor to switch to. We are the market, and the market is now total. Total is my word. I workshopped it. "Seize" scares people. "Total control of their markets" tests at 71, because nobody pictures a kid from Ohio on a beach when you say markets. We moved 22 tankers through Hormuz at night with the lights off. The President said he wanted to announce it so badly it hurt, and he held it a month anyway. That is the most discipline this administration has ever shown. 100 million barrels, he said. The fact-checkers say the math doesn't hold. They're junior. The number was never the deliverable. The number is the soft launch. He said he loves the inflation. People clutched their receipts. 90 a barrel instead of 250 is the discount, and 250 is a price I invented so this one would feel like mercy. How long will we be there? The President answered that. "We had to be there for a while." A while is my favorite contract length. It renews on its own. No court can find the end date. He wondered out loud whether America has the stomach for it. I don't need the country's stomach. I need its signature on page 1. He called the people who object stupid. I'd defend them. They're not stupid. They just read to page 14. Last thing. Annex C. The out-year targets. 3 more, redacted, marked pre-decisional. I can't tell you which 3. Total control of their markets. Nobody in that room asked which market. The crude still sails east, to China, the same as yesterday. We don't own the oil. We own the valve, the fear, and the word "total." That was always the only product.show more

sorry, they just did WHAT someone gave a machine... one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓show more

The thing I help hitters with the most.. (1... good way to train it👇🏼) They start on time but never actually get loaded to GO. They think they’re ready. They’re not. So they start earlier next time. Same problem. —— This is a conversation I have with almost every hitter. When is the right time for you? - Based on your load style - Based on how long your load takes - Based on the pitcher’s velo, windup, or slide step —— If that’s you, try a “loaded at release” type load. Bobby Witt. Judge. Get your hip and elbow loaded early. Then stride with it ready to fire. “The pitcher is up there for 20 seconds telling you to get ready. Shame on you if you’re not ready to swing.” ____ If I have you in the cage, what I’ll do is have you get into a no stride position, get ready to swing and I stand there for 1-4 seconds… Holding my hand out over the plate ready to snap my fingers. When I snap, you swing. Do that 4-5x. You get the quickness, readiness feeling. Then we go back to stride with moving ball and you’re trying to time up that same GO move with the ball With your foot up in the air and the GO will make the foot go down. Not foot down then swing. That typically kills quickness. This works 95% of the time I’d say! —— > Send this to a hitter who needs it > Save it for when you need it.show more

Women are revising their social media boyfriend requirements for... 2026. OUT: 🆇 “Must be over 6 feet.” 🆇 “Must make over 6 figures.” 🆇 “Must own his own home.” IN: ✅ “Must like me.” ✅ “Must not be angry all the time.” ✅ “Must not cheat 24/7.” This is (in part) due to the male retreat from the dating space. There are more long-term single men than at any time in at least a century. Single men are more disinterested in dating and women than any time in at least a century. Women, as a result, are getting approached less, going on fewer dates, and running into far fewer men who actually want relationships (and a lot more guys ‘just playing the field’). You’d probably have to go back to the 1850s, with 300,000 men pushing west to seek their riches in California, living in frontier towns with 10-to-1 male-to-female ratios, to find a time with as many single men unfocused on women. (The shift in women’s ‘going rate’ is also due to the real economy declining… just like how in the roaring 1920s, when the economy was great and everything kept going up, women were flappers who delayed marriage and called wedding rings ‘handcuffs’ and ‘manacles’, but after the stock market crashed in 1929, and the Great Depression got underway, the flappers disappeared and the women of the 1930s suddenly became much more interested in getting hitched.) Anyway: If you are one of the dudes who has foresworn women and dating: thanks for your service, brother 🫡 If you are one of the dudes who is out there getting after it, though: you can look forward to a continuation in 2026 of women softening their stances, making themselves MORE available, asking you to jump through fewer hoops than ever, and acting clingier and needier if they think they have a chance with you. Just make sure you are out IN FIELD chatting up women, flirting with women, getting them onto dates, and inviting them home. You just need to be IN the game. A lot of guys aren’t in the game anymore (or they are locked up on screens, trying to do stuff online, trying to get girls by liking their pics, ‘sliding into their DMs’, where the odds are seriously stacked against ordinary men… ESPECIALLY compared to IRL). 2026 is bringing an increasingly WIDE open field with girls. FOR the guys who’ll be out there playing in it, that is!show more

how to CRACK the tiktok algorithm in 2026... the... algorithm in 2026 is not judging your account. it's judging your first 500 views. everyone overthinks this. here's the actual method to crack it: every video starts from ZERO. doesn't matter if your last one hit 1M or your account has 200k followers. tiktok takes the new post, runs it through content moderation first, THEN drops it to a small cold test pool. that pool isn't random. it's people who already watch your exact niche, because tiktok's whole goal is keeping people on the app, so it shows your video to the people most likely to sit through it. then it reads 4 signals, in this order of weight: completion rate (did they finish it) rewatches (did they loop it) shares (did they send it to someone) velocity (how FAST all of that came in) clear the wave, it pushes ~5x bigger. clear that, 5x again. it keeps expanding until the numbers drop below the bar for that wave, then it stops. that's it. that's "going viral." it's just a video that kept clearing waves. nothing mystical. the part that breaks people's brains: this resets EVERY post. tiktok is not youtube shorts. on youtube the channel carries momentum. on tiktok the VIDEO carries itself. your last banger does not pre-load the next one. all followers do is get you shown FIRST in the test pool, which is a head start, not a guarantee. every video still earns its reach from scratch. now the myths that waste your time: "warm up a new account by scrolling for 2-3 days before posting." MYTH. tiktok has said this directly, there is no warm-up requirement. the only real device-level flag is ban evasion, and that only triggers if your device/IP is tied to a previously banned account. clean device = post day one, no penalty. "posting too much hurts the algorithm." also wrong. there is no volume throttle on your account. LOW volume is the real risk. every post is a fresh roll at the test pool. fewer posts = fewer rolls = fewer chances to hit a wave that clears. the accounts that grow fast post 1-3x a day, not once a week. "you need trending audio to go viral." no. trending audio helps discovery marginally but a strong hook on a silent slideshow will beat a weak video on the #1 sound every time. the sound doesn't save a bad first 2 seconds. what ACTUALLY moves it: the first 2 seconds are the whole game. that IS your completion rate. weak hook, the test pool bounces, the video dies in wave 1 and never recovers. front-load the payoff, cut the intro, no "hey guys." shares > likes, by a lot. a like is passive, it barely registers. a share tells tiktok "spread this" and it's the single signal that widens the wave hardest. build for the share: a take people want to send to someone, a "wait what" fact, a screenshot-able line. velocity beats total. 200 views in the first hour tells the algo more than 2,000 over a week. post when your specific audience is actually awake, check your analytics, don't guess "peak times." the format is disposable, the system isn't. 90% of your posts will die under 5,000 views, most under 300. that's not failure, that's the model working. you're not making one perfect video, you're feeding the test pool enough clean shots that a few clear all the waves. 10 accounts x 2 posts a day = 20 rolls daily. one hits 500k and pays for the other 19. and if monetization is the goal: creator rewards in 2026 is 10k followers + 100k views in the last 30 days, ROLLING (a viral month 6 months ago counts for nothing). only videos 60 SECONDS or longer earn a cent. and you have to be based in an eligible country to even apply: US, UK, germany, france, japan, korea, brazil, a few others. pakistan, india, most of MENA are not on the list. so "target US for higher RPM" isn't just a payout tip, geo is a hard gate on whether the program exists for you at all. stop trying to game it. feed it clean videos with a killer first 2 seconds and let the test pool do the sorting. the algo is dumber and more fair than people think.show more

Here’s my analysis on NotPixel Airdrop: - Mining phase:... it was entertaining to paint on the canvas alongside with many of you! Personally, I had fun drawing 🧡 together with you, just for fun, no competition. For that I will rate it a ✅ - Earn Launchpool: personally, I think this is great that there’s such a tool, that allows holders effortlessly farm tokens. For me it’s a ✅ - Distribution: the major frustrations has been caused by the overload of the smart contract for Airdrop claim, where the biggest frustrations came from the fact that many were not able to claim their tokens for long (Including me) and watched the price go from $0,5 to $0,8 & to $0,2. Personally, I think that if Airdrop claim was done prior to the launch, it may have been better. For that, I will put a ❌ - Launch: NotPixel decided to launch its PX token only on DEX’es on TON. the price at launch actually was at $0,5-$0,8 however, it shortly went down to $0,2 , causing frustration to the community & making them question “who sold?!” & “why no CEX listings?” Here, it’s important to understand the way Sasha & his team decided that they want to build the project. As they mentioned on X “start from the basics, start from 0.” While many dislike this approach, I personally believe that this is how the launches must be happening. Launch on DEX -> generate traction & volume -> go list on CEX’es -> create more news -> grow more from new ppl joining For that I will put ✅ - Team allocation: while many have speculated that “team sold”. this is completely not true. In fact, team has locked their allocation & it can be verified on chain. Which is ✅ for me. - Future plans: NotPixel will have a big canvas on which you will be able to paint, but this time, with some NFT mechanics & PX token integration. For which it is ✅ to me. Burning: NotPixel plan with the new update also includes a mechanism for Burning, which means that PX token will be deflationary. Which is ✅ - ownership revoked: meaning that PX token supply cannot be increased. Nor any modifications could be made to the smart contract at this point. Which is a big ✅ in my opinion - Community: while many are still frustrated with the current token price on the market, many express desire for token appreciation & further growth. One of the interesting feedbacks I saw on the timeline was “NotPixel forced us all to hold”. Which shows that there are ppl who are not willing to get less than their target. Also many express that they want to see PX on CEX’es. For it to grow even more. ✅ - Holders: as with any Airdrop, many have sold their tokens on the market, the volume also came down a bit. However, what’s interesting is that while ~200k ppl sold, the token still fluctuates between the lowest point $0,13 to $0,26. Which shows that there’s a demand & ppl are accumulating more tokens. Currently PX has 205k holders. Which is a ✅ - BuyBacks: the NotPixel team recently announced the buybacks with the money raised from the NotPixel stickers sale. Which created more support for the price on the market. ✅ - Expectations vs reality: Many community members said that they were expecting price to be above $1 as minimum, many were predicting as high as $10. In reality, the price at launch turned out to be much lower than many expected. Leading to the frustration being expressed on the timeline. Which is understandable. ❌ Conclusion: - Mining phase: ✅ - Earn Launchpool: ✅ - Distribution: ❌ - Launch: ✅ - Team Allocation: ✅ - Future plans: ✅ - Burning: ✅ - Ownership revoked: ✅ - Community: ✅ - Holders: ✅ - BuyBacks: ✅ - Expectations vs reality: ❌ This is not bad execution, not many ppl will be able to pull this off in crypto. No one knows the future, but I chose to believe that Sasha & NotPixel team has a plan. Currently they are not listed on any CEX’es, but I think that Sasha & his team could easily get there. Let’s see what the future holds for us 🙌show more

🦕💬 • 260611 [01:03 AM KST] 41 voice notes... transcripted/translated in order: hi how did you like made by riize? you’re asking if i ate dinner? i didn’t eat. i ate lunch kind of late, but i ate a hearty lunch. i also wanted to do a live as i finished work, but while wondering if i should do a live or not, something new to do popped up. and while i was doing it, it got too late. so i'll do a live later. go knicks! lately, the NBA's a hot topic lately. knicks. crazy. i don't know, you know... basketball that well, but i remember growing up they were sort of *the team* that people, yk cheered on. but i don't remember anyone saying that they were doing great, or anything. but apparently they're in the finals rn and they're doing rly well, so also can't forget the world cup. world cup season rn, so... just sports season.. sports season, crazy, yeah? [singing dyd] how is it? riize's challenge. easy, right? do it lots, please~ these days, a lot of short-form [videos] are coming up, as you've probably seen. I thought it might be fun if everyone did that together... what do you think? live... there's a way to turn it on, but even if i turned it on, i'm not sure if i could do it well. i don't have confidence rn so i'm talking like this wa, but there's only 4 days left for our comeback.. only 4 days left... how is it? are you guys looking forward to it? are u guys excited? we're going to ulsan the 15th, and wonbin hyung, said he's buying us lunch... not really. but it's our first time going to ulsan all together. we'll have our comeback there. i think it'll be fun. also, the 15th we're doing a comeback live.. we've been doing it every comeback... and at the live stream we'll be high school students.. but it could honestly be university students too.. or really just do it without a concept, but i think it'll be fun... just... it'll be fun... so, while doing that we'll talk about the comeback whatever happens i want us all to just create fun memories together, and since it’s summer, how about you all enjoy summer together with riize... how does that feel? how's this pfp? i like it ㅎㅎ engddongi, eng, engddon ah, right. also with taro hyung, we'll appear at salon drip... i also think that will be fun when it comes out. the recording was also fun, with doyeon nuna.. please watch it ah... you guys can hear the dryer? i won't tell you who's clothes it's drying, it's a secret~ since you could look at this as something im trying out instead of doing a live stream i won't be able to do it super often or every single day, but I think it wouldn't be bad to use this feature again sometime what do you think? ah, but languages are difficult it's not easy to unify into just one language now, but on the other hand, it's really fun. somehow, this chat room..? is that right? realizing that this place is totally global and international is in a way pretty cool. haha anyway, after a long time i rode a bike again.. there was a lot of sunlight... sunlight.. did lots of photosynthesis but what was a real mistake was first of all the bicycle didnt have a basket, and I didnt wear a comfortable backpack. i came out carrying a somewhat awkward/uncomfortable bag, so i really messed up on top of that the bicycle was incredibly heavy, but then the battery completely died... so I almost cried. without even realizing it I had gone so far away so I suffered a bit. that’s the story i also have the feeling i bruised my butt a bit... ㅎㅎ but that will also become a funny memory, right? how's riizing summer vacation? i think there were really a lot of funny moments... sungchan hyung was so funny... sungchan hyung said somewhere he hasn't eaten ramyeon in a year ㅎㅎ and sungchan hyung honestly isn't the type to eat ramen alone, he's on the not eating it side... but he's always like 'one bite, one bite'... so if we put all that together, wouldn't it be a plate of ramen?ㅎㅎ for me, at the beginning of our debut, it was a bit... you know how while filming content, i ended up grilling meat and all? but now, grilling meat... well, i don't think it's that bad also, back when we went to japan to shoot our album jacket photos and the trailer, we visited a lot of vinyl (LP) shops... and that was actually the first time in my life buying an LP with my own money when i was young i went to la to play, and then there was this famous 'amoeba music' store. i looked around and everything, but honestly back then i really wasn't that much interested in LPs. so i bought things like a lot of stickers and used them to decorate my laptop. somehow, after that time, [it was like] I truly spent time in a vinyl shop? but visiting now that I'm a bit older, it was so much fun there are so many songs I don’t know, and tons of songs that aren't even famous so back then in japan, duran duran is a really famous band, but I bought duran duran's Notorious LP and then I also bought Barry Finnertys,, barrynims NY City LP. what was so surprising was that as soon as I listened to the LP, the sound and quality were so good that I was like, 'wow, what is this?' but when I looked it up on the internet, it only had like 3,000 views on YouTube? I think it was 3,000, so because it was so good, I bought it. and after that yesterday too while wandering around here and there, i happened to walk into a cafe and they had LPs there too so yesterday, I bought one or two ㅎㅎ lp's no mater how you look it they have a vibe... its true that nowadays you can just look up and listen to everything on your phone, but another good thing about it (lps) is that I think it’s really great for finding songs you don't know so when going to the lp shop you get to learn a lot of new songs that you didn't know before. even if you don't go with the sole purpose of buying an LP, it's just... really great for discovering things... should I say? a riize lp would also be good... back then, during boom boom bass, we even went there and did all that stuff with our LP too... how was the video editing? back then when i went to school when doing things like presentations, together with other students we would edit videos.. i did lowkey find that kind of stuff pretty fun though because it was pretty much my first time [doing that] since back during those days... but really i think it was fun like... yeah.. i want to try to do it again when i was young i did draw. i drew for fun but it had been such a long time since i last properly sat down and tried to draw something properly with a pencil.. i think it's been such a long time i'm probably the worst at drawing in my family? ㅎㅎ my dad's good, my mom's also good.. my grandpa also draws reaññu well. but it was fun drawing again, so i also want to try that again... why is there so many things i want to do? really why is it? and above all, i want to work really hard on music so that someday i can show you even more. now, like from the trailer to the full version that was uploaded to our site... those kinds of things are a bit different from just being a song, though.. now everyone, let's try meditating together with me sorry.. anyways.. i have to leave in a couple hours actually, so.. time to hit the hay, yeehaw i can't fall asleep right away ㅎㅎ this.. seems like it might be a bit difficult to interpret/translate.. (yes anton, it is,,,) next time it might be better to just turn on a live stream instead ㅎㅎ sorry wow... sleep is really not coming... [1 video] [1 picture] ok jinja gnshow more