Loading video...

Video Failed to Load

Go Home

LFM2-8B-A1B just dropped on Hugging Face! 8.3B params with only 1.5B active/token 🚀 > Quality ≈ 3–4B dense, yet faster than Qwen3-1.7B > MoE designed to run on phones/laptops (llama.cpp / vLLM) > Pre-trained on 12T tokens → strong math/code/IF

maximelabonne's profile picture

Maxime Labonne

24,701 subscribers

40,714 views • 9 months ago •via X (Twitter)

Science & Technology
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Qwen3.5-35B with only 3B active parameters This MoE model runs FASTER than most 7B dense models. Tested on 3 generations of NVIDIA: - 5090: 137 tok/s - 4090: 112 tok/s - 3090: 78 tok/s The surprise? The 4090 <> 5090 gap is only 22%. With a 3B active MoE, even old GPUs fly.
0:30

Qwen3.5-35B with only 3B active parameters This MoE model runs FASTER than most 7B dense models. Tested on 3 generations of NVIDIA: - 5090: 137 tok/s - 4090: 112 tok/s - 3090: 78 tok/s The surprise? The 4090 <> 5090 gap is only 22%. With a 3B active MoE, even old GPUs fly.

stevibe

69,527 views • 4 months ago

Qwen3-Next (thinking & non-thinking) are now live in BF16 at Hyperbolic! Qwen3-Next is a huge efficiency leap: - 80B MoE with just 3B active params - 10x cheaper to train vs Qwen3-32B - 10x inference throughput for >32K tokens Proud to be a launch partner with Qwen - kudos to this amazing team for keeping pushing open-source AI forward. We’re the first to serve Qwen3-Next on Hugging Face. Give it a try!
0:23

Qwen3-Next (thinking & non-thinking) are now live in BF16 at Hyperbolic! Qwen3-Next is a huge efficiency leap: - 80B MoE with just 3B active params - 10x cheaper to train vs Qwen3-32B - 10x inference throughput for >32K tokens Proud to be a launch partner with Qwen - kudos to this amazing team for keeping pushing open-source AI forward. We’re the first to serve Qwen3-Next on Hugging Face. Give it a try!

Yuchen Jin

84,342 views • 10 months ago

Introducing Lucy: 1.7B model that Google for you It's an agentic‑search model that can even run on your phone. - Agentic search on tap - Lucy calls tools ( ‑aware) - Fits in your pocket - runs on CPU or mobile Under the hood: - Built on Qwen's Qwen3‑1.7B - Smooth multi‑category rewards replace brittle if‑else scoring - Task‑vector RLVR optimizes the "thinking" tag for targeted search moves. Benchmarks: - SimpleQA + MCP = 78.3 - Close to Jan‑Nano-4B (80.7) Run locally: - Demo uses vLLM in Jan - You can spin it up with Georgi Gerganov's llama.cpp or vLLM Models on Hugging Face: - Lucy 1.7B 40k: - Lucy 1.7B 128K:
0:33

Introducing Lucy: 1.7B model that Google for you It's an agentic‑search model that can even run on your phone. - Agentic search on tap - Lucy calls tools ( ‑aware) - Fits in your pocket - runs on CPU or mobile Under the hood: - Built on Qwen's Qwen3‑1.7B - Smooth multi‑category rewards replace brittle if‑else scoring - Task‑vector RLVR optimizes the "thinking" tag for targeted search moves. Benchmarks: - SimpleQA + MCP = 78.3 - Close to Jan‑Nano-4B (80.7) Run locally: - Demo uses vLLM in Jan - You can spin it up with Georgi Gerganov's llama.cpp or vLLM Models on Hugging Face: - Lucy 1.7B 40k: - Lucy 1.7B 128K:

Menlo Research

20,205 views • 1 year ago

We've partnered to bring more Gemma 3 quantized models to you! 🚀 We worked with Georgi Gerganov llama.cpp, LM Studio, MLX, ollama to make sure you can run it using your favorite tool! Gemma models optimized with QAT, reduce memory requirements while keeping quality! All models checkpoints are available on Hugging Face and Kaggle. 🤗 What does this mean? - Gemma 3 27B (int4): Fits on NVIDIA RTX 3090 (24GB VRAM) or similar. - Gemma 3 12B (int4): Only needs a NVIDIA RTX 4060 (8GB VRAM) or similar. - Gemma 3 4B, 1B (int4): Run anything with more than 2.5GB Memory. Want to see it in action? Video below shows how easy it is to get started using LMStudio:
0:20

We've partnered to bring more Gemma 3 quantized models to you! 🚀 We worked with Georgi Gerganov llama.cpp, LM Studio, MLX, ollama to make sure you can run it using your favorite tool! Gemma models optimized with QAT, reduce memory requirements while keeping quality! All models checkpoints are available on Hugging Face and Kaggle. 🤗 What does this mean? - Gemma 3 27B (int4): Fits on NVIDIA RTX 3090 (24GB VRAM) or similar. - Gemma 3 12B (int4): Only needs a NVIDIA RTX 4060 (8GB VRAM) or similar. - Gemma 3 4B, 1B (int4): Run anything with more than 2.5GB Memory. Want to see it in action? Video below shows how easy it is to get started using LMStudio:

Philipp Schmid

15,996 views • 1 year ago

🎂 Cerebras Inference turns 1! 🚀 Let's break it down: - 6x faster than when we launched — From Meta Llama to Qwen 3 to OpenAI OSS, models running on Cerebras deliver 𝟯,𝟬𝟬𝟬+ 𝘁𝗼𝗸𝗲𝗻𝘀/𝘀𝗲𝗰 - Our inference powers the best AI natives, global enterprises, and developers including Meta, IBM , AlphaSense, Docker, GSK , Mayo Clinic, Core42, Vercel and more - Largest model served: ~half a trillion parameters, 7x bigger than launch - #1 provider of tokens on Hugging Face 🤗 - Serving billions of tokens per day on OpenRouter - The leading code gen inference provider, powering AI developers with Cline and Windsurf We’re just getting started… but couldn’t have done Year 1 without YOU. Build on, builders. ❤️‍🔥
0:39

🎂 Cerebras Inference turns 1! 🚀 Let's break it down: - 6x faster than when we launched — From Meta Llama to Qwen 3 to OpenAI OSS, models running on Cerebras deliver 𝟯,𝟬𝟬𝟬+ 𝘁𝗼𝗸𝗲𝗻𝘀/𝘀𝗲𝗰 - Our inference powers the best AI natives, global enterprises, and developers including Meta, IBM , AlphaSense, Docker, GSK , Mayo Clinic, Core42, Vercel and more - Largest model served: ~half a trillion parameters, 7x bigger than launch - #1 provider of tokens on Hugging Face 🤗 - Serving billions of tokens per day on OpenRouter - The leading code gen inference provider, powering AI developers with Cline and Windsurf We’re just getting started… but couldn’t have done Year 1 without YOU. Build on, builders. ❤️‍🔥

Cerebras

34,546 views • 11 months ago

Auto regressive LLMs are officially on notice. run Gemma 4 26B diffusion gguf with llama.cpp Google just dropped DiffusionGemma-26B, and it completely flips how we generate text. instead of predicting words one by one, it generates 256 tokens in parallel using bi-directional attention. its like stable diffusion, but for language. the model starts with random text "noise" and iteratively refines and self-corrects the entire block in real-time to fix formatting and reasoning errors on the fly. since it’s a Mixture of Experts (MoE) that only activates 3.8B parameters during inference, it fits perfectly on consumer hardware. You can run the Q4_K_M quant with an 18GB VRAM budget on a single RTX 3090 or RTX 4090 with exceptional throughput. Tested on Ubuntu 22 with CUDA 13.1 using the cutting edge experimental llama.cpp branch. Here is how to compile and run it with the live terminal denoising visualizer: # 1. Clone & check out the experimental PR (#24423) - 1) git clone && cd llama.cpp -git fetch origin 2) pull/24423/head:diffusiongemma && --git checkout diffusiongemma # 2. Build with CUDA support 1) cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native 2) cmake --build build -j $(nproc) --config Release --target llama-diffusion-cli # 3. Run with live visual denoising (llama.cpp flags) ./build/bin/llama-diffusion-cli \ -m /path/to/diffusiongemma-26B-A4B-it-Q4_K_M.gguf \ -ngl 99 -cnv -n 2048 --diffusion-visual Watch the video below to see the live --diffusion-visual canvas iteratively de noising the prompt output in real time. guide and unsloth's hugging face GGUF model links are in the comments below! Is auto regressive generation officially legacy tech? Let me know what you think.
0:31

Auto regressive LLMs are officially on notice. run Gemma 4 26B diffusion gguf with llama.cpp Google just dropped DiffusionGemma-26B, and it completely flips how we generate text. instead of predicting words one by one, it generates 256 tokens in parallel using bi-directional attention. its like stable diffusion, but for language. the model starts with random text "noise" and iteratively refines and self-corrects the entire block in real-time to fix formatting and reasoning errors on the fly. since it’s a Mixture of Experts (MoE) that only activates 3.8B parameters during inference, it fits perfectly on consumer hardware. You can run the Q4_K_M quant with an 18GB VRAM budget on a single RTX 3090 or RTX 4090 with exceptional throughput. Tested on Ubuntu 22 with CUDA 13.1 using the cutting edge experimental llama.cpp branch. Here is how to compile and run it with the live terminal denoising visualizer: # 1. Clone & check out the experimental PR (#24423) - 1) git clone && cd llama.cpp -git fetch origin 2) pull/24423/head:diffusiongemma && --git checkout diffusiongemma # 2. Build with CUDA support 1) cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native 2) cmake --build build -j $(nproc) --config Release --target llama-diffusion-cli # 3. Run with live visual denoising (llama.cpp flags) ./build/bin/llama-diffusion-cli \ -m /path/to/diffusiongemma-26B-A4B-it-Q4_K_M.gguf \ -ngl 99 -cnv -n 2048 --diffusion-visual Watch the video below to see the live --diffusion-visual canvas iteratively de noising the prompt output in real time. guide and unsloth's hugging face GGUF model links are in the comments below! Is auto regressive generation officially legacy tech? Let me know what you think.

Alok

52,656 views • 1 month ago

Today we’re releasing Chroma 1.0 → the world first open-source, end-to-end, real-time speech-to-speech model → with personalized voice cloning Trained by FlashLabs. Deployed on FlashAI👉 An open research-grade alternative to the OpenAI Realtime model. Voice Test dubbing Elon Musk and Lex Fridman: 🔥What’s real (evals and benchmarks attached): ⚡ 2× faster than real-time ) 📘 SGLang Cookbook: 📄 Paper + benchmarks: 🤗 Models: 💻 Inference code: 🔁 RT if you believe real-time AI should be open 💬 Reply if you’re building on Conversational Voice AI products
1:57

Today we’re releasing Chroma 1.0 → the world first open-source, end-to-end, real-time speech-to-speech model → with personalized voice cloning Trained by FlashLabs. Deployed on FlashAI👉 An open research-grade alternative to the OpenAI Realtime model. Voice Test dubbing Elon Musk and Lex Fridman: 🔥What’s real (evals and benchmarks attached): ⚡ 2× faster than real-time ) 📘 SGLang Cookbook: 📄 Paper + benchmarks: 🤗 Models: 💻 Inference code: 🔁 RT if you believe real-time AI should be open 💬 Reply if you’re building on Conversational Voice AI products

FlashLabs

1,248,697 views • 6 months ago

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.
2:02

first impressions of qwen 3.5 27B dense on a single RTX 3090. 35 tok/s. from 4K all the way to 300K+ context. no speed drop. hermes 4.3 started at 35 and degraded to 15 as context filled. qwen dense holds. MoE held 112 flat. 3x faster but only 3B of 35B active per token. architecture tradeoff. Q4_K_M on 16.7GB. native context 262K. pushed past training limit to 376K before VRAM ceiling on 24GB. tried q8 KV cache at 262K, speed collapsed to 11 tok/s. q4_0 KV is the sweet spot. flash attention mandatory. built in reasoning mode. the model thinks step by step before it answers. full chain of thought surviving Q4 quant. 1,799+ token thinking chains with self correction loops. on a single consumer GPU. gave it one prompt: "build a realtime particle galaxy simulation in one HTML file." 3,340 tokens. 95 seconds. one shot. ran on first load. full reasoning and coding in the video below. optimal config if you want to skip the hours of testing: llama-server -ngl 99 -c 262144 -fa on --cache-type-k q4_0 --cache-type-v q4_0 this is just the warmup. octopus invaders is next: 10 files, 3,400+ lines, zero steering. the prompt hermes quit at 22%. already more impressed than expected. full results coming soon.

Sudo su

120,788 views • 4 months ago

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.
0:25

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.

Harman

41,179 views • 19 days ago

six months ago this wasn't happening on 8gb vram. running unsloth's Q4_K_XL quant of gemma 4 26b-a4b-it-qat, a sparse MoE model with only 4b active params on a single rtx 4060 laptop gpu, 8gb vram, 20+ tok/s decode. no cloud, no api, no offload hacks. just a gaming laptop on battery. what makes it fit: google's QAT (quantization aware training), plus MTP (multi token prediction) support in the latest llama.cpp builds. that combo is the single biggest unlock for local inference on low vram. rtx 3060, rtx 3070, gtx 1070, gtx 1080, rtx 4050, rtx 4060, rtx 5050, rtx 5060 — any 6-8gb consumer gpu, old or new — this model runs on it. world cup season, so i told it to build a soccer themed flappy bird clone. one shot, zero iteration, fully playable. six months ago an 8gb model could barely clone vanilla flappy bird. now it's shipping a themed game from a sparse MoE model running locally on a laptop battery. inference benchmarks: - decode throughput: 30 tok/s - context: 64k. this is the real unlock. 64k ctx is what makes a hermes agent loop viable locally on this model, not just single-turn chat. llama.cpp flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 64000 -cmoe --port 8080 game's deployed on my own site, built and shipped end to end with open source llm, zero closed source api dependency in the pipeline. link in the description. gguf weights on huggingface, link in the comments. pull it down, run it on whatever 8gb card is sitting in your rig. try the game and tell me your score and what you want in v2. local llms on consumer gpus stopped being a meme.
0:20

six months ago this wasn't happening on 8gb vram. running unsloth's Q4_K_XL quant of gemma 4 26b-a4b-it-qat, a sparse MoE model with only 4b active params on a single rtx 4060 laptop gpu, 8gb vram, 20+ tok/s decode. no cloud, no api, no offload hacks. just a gaming laptop on battery. what makes it fit: google's QAT (quantization aware training), plus MTP (multi token prediction) support in the latest llama.cpp builds. that combo is the single biggest unlock for local inference on low vram. rtx 3060, rtx 3070, gtx 1070, gtx 1080, rtx 4050, rtx 4060, rtx 5050, rtx 5060 — any 6-8gb consumer gpu, old or new — this model runs on it. world cup season, so i told it to build a soccer themed flappy bird clone. one shot, zero iteration, fully playable. six months ago an 8gb model could barely clone vanilla flappy bird. now it's shipping a themed game from a sparse MoE model running locally on a laptop battery. inference benchmarks: - decode throughput: 30 tok/s - context: 64k. this is the real unlock. 64k ctx is what makes a hermes agent loop viable locally on this model, not just single-turn chat. llama.cpp flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 64000 -cmoe --port 8080 game's deployed on my own site, built and shipped end to end with open source llm, zero closed source api dependency in the pipeline. link in the description. gguf weights on huggingface, link in the comments. pull it down, run it on whatever 8gb card is sitting in your rig. try the game and tell me your score and what you want in v2. local llms on consumer gpus stopped being a meme.

Alok

60,866 views • 1 month ago

Most speculative decoding still drafts tokens one at a time. That's not parallel generation — it just hides the serial loop behind a smaller model. UC San Diego's z-lab just drew a clear line between the two. They released DFlash — a lightweight block diffusion model that drafts a whole block of tokens in a single forward pass, then lets the target model verify the block in parallel. Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell. No token-by-token drafting anywhere in the speculative path. Here's what's actually interesting: → The drafter is conditioned on the target model's own hidden features, injected into the Key/Value cache of every draft layer — so acceptance length scales with draft depth instead of diluting away → A 5-layer drafter replaces the 7B diffusion drafters that capped earlier methods near 3–4× → MATH-500 speedup: 6.08× vs. 1.81× for EAGLE-3 (4.86× average vs. 1.76×, Qwen3-8B, greedy) → Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell — at the same interactivity target → Lossless: the target still verifies every token, so output quality is preserved Full analysis: Paper: NVIDIA's metrics: Project: Model weights: Repo: NVIDIA AI NVIDIA AIDev
0:20

Most speculative decoding still drafts tokens one at a time. That's not parallel generation — it just hides the serial loop behind a smaller model. UC San Diego's z-lab just drew a clear line between the two. They released DFlash — a lightweight block diffusion model that drafts a whole block of tokens in a single forward pass, then lets the target model verify the block in parallel. Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell. No token-by-token drafting anywhere in the speculative path. Here's what's actually interesting: → The drafter is conditioned on the target model's own hidden features, injected into the Key/Value cache of every draft layer — so acceptance length scales with draft depth instead of diluting away → A 5-layer drafter replaces the 7B diffusion drafters that capped earlier methods near 3–4× → MATH-500 speedup: 6.08× vs. 1.81× for EAGLE-3 (4.86× average vs. 1.76×, Qwen3-8B, greedy) → Up to 15× higher throughput for gpt-oss-120b on NVIDIA Blackwell — at the same interactivity target → Lossless: the target still verifies every token, so output quality is preserved Full analysis: Paper: NVIDIA's metrics: Project: Model weights: Repo: NVIDIA AI NVIDIA AIDev

Marktechpost AI

23,328 views • 1 month ago

💥 A 450M model just beat bigger VLAs on real robot tasks, and it’s 100% open source [📍 bookmark for later] Came across SmolVLA, a new vision-language-action model for robotics that’s compact, fast, and trained entirely on open community datasets from LeRobot via Hugging Face. What stood out to me is how it matches or outperforms much larger models like ACT using noisy, real-world community data instead of giant private datasets. Why it’s worth a look ✅ 26% performance boost from pretraining on open-source data ✅ Runs on consumer hardware, even a MacBook ✅ 30% faster responses with async inference and smart architecture tweaks ✅ Strong results across Meta-World, LIBERO, SO100, and SO101 ✅ Fully open source: weights, code, training pipeline, eval stack They also introduced smart efficiency tricks like using fewer visual tokens, pulling outputs from mid-layer, and separating perception from action to make it all run fast. SmolVLA is a strong case for what can happen when the robotics community shares data and builds in the open. Definitely worth keeping an eye on.
0:21

💥 A 450M model just beat bigger VLAs on real robot tasks, and it’s 100% open source [📍 bookmark for later] Came across SmolVLA, a new vision-language-action model for robotics that’s compact, fast, and trained entirely on open community datasets from LeRobot via Hugging Face. What stood out to me is how it matches or outperforms much larger models like ACT using noisy, real-world community data instead of giant private datasets. Why it’s worth a look ✅ 26% performance boost from pretraining on open-source data ✅ Runs on consumer hardware, even a MacBook ✅ 30% faster responses with async inference and smart architecture tweaks ✅ Strong results across Meta-World, LIBERO, SO100, and SO101 ✅ Fully open source: weights, code, training pipeline, eval stack They also introduced smart efficiency tricks like using fewer visual tokens, pulling outputs from mid-layer, and separating perception from action to make it all run fast. SmolVLA is a strong case for what can happen when the robotics community shares data and builds in the open. Definitely worth keeping an eye on.

Ilir Aliu - eu/acc

17,353 views • 11 months ago

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?
0:26

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 views • 2 months ago

Ever wished we had fewer X-training hyphenates? Pre, mid, post etc. Why not just Training? Trying to bridge the divides (and get all our friends into one team again), we intro *Introspective X Training*, an offline RL inspired method that scales effectively across any LLM stage by annotating your data with a thinking reward generated language critique! Up to 2.8x FLOP efficiency + 5-10 point score gains (esp with math and code) at any stage from scratch to 24T tokens on 8b (active) sized models!! We burned much compute ablating so you wouldn't have to Moral of the story is‼️don't throw out any data via filtering, just feedback condition it‼️ You can spend FLOPs up front on inference to *classify* data quality and then train so that tokens aren't all treated equally based on the feedback starting early in training itself. Right now they're really only separated out much later during mid/post training This improves overall compute efficiency and gives us benchmark perf not possible with just baseline methods! Paper here: Thanks to Brandon Cui and Ximing Lu for leading this w/ Syeda Nahida Akter David Acuna Hyunwoo Kim Jaehun Jung Yuxiao Qu Shrimai Yejin Choi
0:34

Ever wished we had fewer X-training hyphenates? Pre, mid, post etc. Why not just Training? Trying to bridge the divides (and get all our friends into one team again), we intro *Introspective X Training*, an offline RL inspired method that scales effectively across any LLM stage by annotating your data with a thinking reward generated language critique! Up to 2.8x FLOP efficiency + 5-10 point score gains (esp with math and code) at any stage from scratch to 24T tokens on 8b (active) sized models!! We burned much compute ablating so you wouldn't have to Moral of the story is‼️don't throw out any data via filtering, just feedback condition it‼️ You can spend FLOPs up front on inference to *classify* data quality and then train so that tokens aren't all treated equally based on the feedback starting early in training itself. Right now they're really only separated out much later during mid/post training This improves overall compute efficiency and gives us benchmark perf not possible with just baseline methods! Paper here: Thanks to Brandon Cui and Ximing Lu for leading this w/ Syeda Nahida Akter David Acuna Hyunwoo Kim Jaehun Jung Yuxiao Qu Shrimai Yejin Choi

Prithviraj (Raj) Ammanabrolu

27,471 views • 2 months ago

HERMES AGENT NOW RUNS A FULL BUSINESS FOR $8/MONTH. CONTENT. CODE. INBOX. ALL AGENTS. EVERYTHING WHILE YOU SLEEP. one user built this in one session. zero terminal commands. browser only. the team: → content agent: finds video topics, writes first drafts, helps with animations → code builder: writes website and SaaS code → ops agent: organizes and answers email total setup cost: $1. monthly runtime: $8 VPS + cheap model API. the trick is matching model to task. frontier models (Opus 4.8, GPT-5.5) for complex /goals only. $5/million tokens. mid-tier (Claude Sonnet, Haiku, Kimiko 2.6) for daily content and research work. budget (DeepSeek V4 Flash, MiniMax 2.7, Gemma 4, Qwen 3) for cron jobs and routine. run the initial setup on Opus. switch to DeepSeek for everything after. your bill drops 90% with zero quality loss on routine work. the agent team picks up tasks from kanban. each completed task creates a skill. week 3 the same workflows run faster than week 1. full breakdown on Hermes Os in the article 👇
14:53

HERMES AGENT NOW RUNS A FULL BUSINESS FOR $8/MONTH. CONTENT. CODE. INBOX. ALL AGENTS. EVERYTHING WHILE YOU SLEEP. one user built this in one session. zero terminal commands. browser only. the team: → content agent: finds video topics, writes first drafts, helps with animations → code builder: writes website and SaaS code → ops agent: organizes and answers email total setup cost: $1. monthly runtime: $8 VPS + cheap model API. the trick is matching model to task. frontier models (Opus 4.8, GPT-5.5) for complex /goals only. $5/million tokens. mid-tier (Claude Sonnet, Haiku, Kimiko 2.6) for daily content and research work. budget (DeepSeek V4 Flash, MiniMax 2.7, Gemma 4, Qwen 3) for cron jobs and routine. run the initial setup on Opus. switch to DeepSeek for everything after. your bill drops 90% with zero quality loss on routine work. the agent team picks up tasks from kanban. each completed task creates a skill. week 3 the same workflows run faster than week 1. full breakdown on Hermes Os in the article 👇

YanXbt

30,910 views • 1 month ago

🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠 Hybrid Linear–Softmax Attention 3 gated linear layers + 1 gated-softmax anchor (75% linear / 25% softmax) 🧱 Block Attention Residuals (AttnRes) Boosts deep-layer effective rank by ~12% 🏗️ Unified 5B & 14B models Trained from scratch under very limited resources: • 5B → only 16 nodes of H100 • 14B → 48 nodes of B200 Results 📊 • 84.30 VBench Total • 3.2× faster DiT forward than matched full-softmax at 720p/60s • 720p/5s in 13.06s on a single H100 with Sol-Engine • 120× faster than Wan 2.2-A14B under the same one-H100 setup SANA-Video 2.0 shows that high-quality 720p video generation can be both efficient and practical on a single GPU. 🎬 Project: 📄 Paper: 💻 Code: Proud of the team! 🎉 More details below 🧵
0:24

🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠 Hybrid Linear–Softmax Attention 3 gated linear layers + 1 gated-softmax anchor (75% linear / 25% softmax) 🧱 Block Attention Residuals (AttnRes) Boosts deep-layer effective rank by ~12% 🏗️ Unified 5B & 14B models Trained from scratch under very limited resources: • 5B → only 16 nodes of H100 • 14B → 48 nodes of B200 Results 📊 • 84.30 VBench Total • 3.2× faster DiT forward than matched full-softmax at 720p/60s • 720p/5s in 13.06s on a single H100 with Sol-Engine • 120× faster than Wan 2.2-A14B under the same one-H100 setup SANA-Video 2.0 shows that high-quality 720p video generation can be both efficient and practical on a single GPU. 🎬 Project: 📄 Paper: 💻 Code: Proud of the team! 🎉 More details below 🧵

Enze Xie

23,585 views • 3 days ago

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own
0:20

HERMES AGENT NOW RUNS ON AN 8GB LAPTOP GPU JUST AS EASILY AS IT RUNS ON A 128GB MINI PC Nous Research shipped the official Hermes Agent Desktop App this week. Someone pointed it at a local llama server running on an RTX 4060 with 16GB system RAM. The integration took two minutes The model behind it: Gemma 4 26B MoE, QAT quantized, running on 8GB of VRAM. A 60k token prompt held a stable 20 tokens a second, flat, no slowdown as context grew. The flags were nothing exotic, just -cmoe -c 248000 on llama.cpp What that 8GB setup does out of the box: reads and patches its own code, runs it in a terminal, debugs errors, manages GitHub repos, spawns sub-agents for parallel work. Browses the web with vision to debug a UI. Schedules cron jobs in plain language. Connects to Notion, Google Workspace, Linear, and Obsidian to manage tasks on its own That's the same agent layer running on a Minisforum MS-S1 MAX with 128GB of unified memory, 96GB of it to the GPU, holding a 120B model at 56 tokens a second instead of a 26B model at 20. Same software, same tool execution, same zero API key. The only thing that changes between an $800 laptop and a $2,000 mini PC is how big a model you can afford to run underneath it The barrier to running a real autonomous agent locally didn't just drop. It dropped all the way down to hardware most people already own

NO1ennn

40,079 views • 1 month ago

Alibaba just dropped Qwen3.5-397B-A17B and there's a lot to unpack. 397B params, 17B active per forward pass. Sparse MoE done right. But the real story isn't the size—it's the architecture choices. The MoE Design Most MoE models feel like bolt-ons. Qwen 3.5's sparse activation is native—only 4.3% of parameters fire per token. That's how you get trillion-parameter-class performance without trillion-parameter inference costs. The 0.8 RMB/million tokens pricing isn't subsidized; it's structurally earned. Native Multimodal, Not Glued-On This is a vision-language model from the ground up. Heterogeneous architecture—separate processing pipelines for text, image, video that fuse early. Not a vision encoder slapped onto an LLM. The result: 90.8 on OmniDocBench, 79.0 on MMMU-Pro. Document understanding and visual reasoning without the usual brittleness. The Context Window Reality Qwen3.5-Plus (the hosted version) ships with 1M tokens by default. That's not a marketing number—they're actually positioning it for long-document workflows. With built-in adaptive tool use, it's clearly aimed at agentic automation, not just chat. What Actually Impressed Me • FP8 native pipeline: ~50% activation memory reduction • Async RL framework for continuous refinement—training and inference workloads separated • 201 languages (up from 119), 250k vocab for better low-resource encoding • Apache 2.0 license. Full weights on HuggingFace and ModelScope. The Benchmark Context 76.4 on SWE-bench Verified puts it in the range where it can handle real debugging workflows. 72.9 on BFCL v4 for agentic tool use. 88.4 on GPQA Diamond. These aren't SOTA in isolation, but the breadth is unusual—strong across reasoning, coding, multimodal, and agentic tasks. The Honest Caveat I haven't stress-tested the 1M context for needle-in-haystack retrieval yet. And "native multimodal" claims need real-world torture testing—PDFs with tables, charts, mixed layouts. Benchmarks are benchmarks. Bottom Line This isn't just another model release. It's a bet on efficient scale: big model capabilities, small active compute, open weights. At 1/18th the cost of Gemini 3 Pro, it's going to force pricing conversations across the board.
1:03

Alibaba just dropped Qwen3.5-397B-A17B and there's a lot to unpack. 397B params, 17B active per forward pass. Sparse MoE done right. But the real story isn't the size—it's the architecture choices. The MoE Design Most MoE models feel like bolt-ons. Qwen 3.5's sparse activation is native—only 4.3% of parameters fire per token. That's how you get trillion-parameter-class performance without trillion-parameter inference costs. The 0.8 RMB/million tokens pricing isn't subsidized; it's structurally earned. Native Multimodal, Not Glued-On This is a vision-language model from the ground up. Heterogeneous architecture—separate processing pipelines for text, image, video that fuse early. Not a vision encoder slapped onto an LLM. The result: 90.8 on OmniDocBench, 79.0 on MMMU-Pro. Document understanding and visual reasoning without the usual brittleness. The Context Window Reality Qwen3.5-Plus (the hosted version) ships with 1M tokens by default. That's not a marketing number—they're actually positioning it for long-document workflows. With built-in adaptive tool use, it's clearly aimed at agentic automation, not just chat. What Actually Impressed Me • FP8 native pipeline: ~50% activation memory reduction • Async RL framework for continuous refinement—training and inference workloads separated • 201 languages (up from 119), 250k vocab for better low-resource encoding • Apache 2.0 license. Full weights on HuggingFace and ModelScope. The Benchmark Context 76.4 on SWE-bench Verified puts it in the range where it can handle real debugging workflows. 72.9 on BFCL v4 for agentic tool use. 88.4 on GPQA Diamond. These aren't SOTA in isolation, but the breadth is unusual—strong across reasoning, coding, multimodal, and agentic tasks. The Honest Caveat I haven't stress-tested the 1M context for needle-in-haystack retrieval yet. And "native multimodal" claims need real-world torture testing—PDFs with tables, charts, mixed layouts. Benchmarks are benchmarks. Bottom Line This isn't just another model release. It's a bet on efficient scale: big model capabilities, small active compute, open weights. At 1/18th the cost of Gemini 3 Pro, it's going to force pricing conversations across the board.

Bo Wang

13,221 views • 5 months ago

A single RTX 4090 (24 GB VRAM) can run the updated gemma 4 31B (dense) model with a 190,000 context window at 33 tokens/second. The VRAM barrier is dying. Google quietly updated Gemma 4, and Unsloth immediately compiled the new quants. I built llama.cpp from source on Ubuntu 22 to benchmark it. Google's stealth update 2 days ago enabled uniform Flash Attention 4 on Hopper to boost prefill and patched the chat template to improve tool calling. The agentic reasoning gains on the benchmark charts are massive: TB2 (Agents): +4.5% (to 25.8%) Tau2 (Telecom): +10.1% (to 62.7%) Running on Ubuntu 22, CUDA 13.0 with a single NVIDIA GeForce RTX 4090. Here is the exact step by step benchmarking process with a massive 28k tokens prompt and the commands I used to squeeze out maximum context without killing my throughput: # 1. The Baseline (Unquantized KV Cache) I started with full GPU offload (-ngl 99) and pushed the context to 40k. llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -ngl 99 -c 40000 -fa on --port 8080 -v VRAM: 23.8 GB (maxed out on card) Throughput: Prefill: 2198.81 t/s | Decode: 35.77 t/s (with 28k tokens prompt) # 2. The CPU Split Trap I tried stretching to 80k context by offloading layers to the CPU (-ngl 52). llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 80000 -ngl 52 -fa on --port 8080 -v Throughput: Prefill: 1212.73 t/s | Decode: 5 t/s (with 28k tokens prompt) # 3. The KV Quantization Breakthrough Instead of spilling layers to the CPU, I kept the model fully on card (-ngl 99) but enabled 8-bit KV cache quantization to free up VRAM. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 100000 --cache-type-k q8_0 --cache-type-v q8_0 -ngl 99 --port 8080 -v VRAM: 23.9 GB Throughput: Prefill: 2139.68 t/s | Decode: 32 t/s (with 28k tokens prompt) Result: 100k tokens of context on a single GPU with practically zero speed loss (and minimal intelligence loss). # 4. The Limit Test (Q4 KV Cache) To find the absolute breaking point, I dropped the KV cache to 4 bit (q4_0) and set -c 190000. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 190000 --cache-type-k q4_0 --cache-type-v q4_0 -ngl 99 --port 8080 -v VRAM: 23.8 GB Throughput: Prefill: 2206.66 t/s | Decode: 33 t/s (with 28k tokens prompt) (Note: Pushing it to 220k required dropping to -ngl 58 again, which immediately penalized decode down to 17 t/s). # The Tradeoff: For Max Reasoning: Keep your KV cache unquantized (f16). You get pristine reasoning but hit a strict 40k context ceiling. For Massive Document Retrieval: If you need to feed the model giant codebases, use --cache-type-k q4_0. Getting 190k context at 33 tokens/second on a consumer desktop with a 31b dense model is a cheat code. If you’re rocking a single 3090 or 4090 and slept on Gemma 4 earlier, this update is your cue to dust off the terminal. Hugging Face links to the Unsloth QAT quants are in the replies below.
0:20

A single RTX 4090 (24 GB VRAM) can run the updated gemma 4 31B (dense) model with a 190,000 context window at 33 tokens/second. The VRAM barrier is dying. Google quietly updated Gemma 4, and Unsloth immediately compiled the new quants. I built llama.cpp from source on Ubuntu 22 to benchmark it. Google's stealth update 2 days ago enabled uniform Flash Attention 4 on Hopper to boost prefill and patched the chat template to improve tool calling. The agentic reasoning gains on the benchmark charts are massive: TB2 (Agents): +4.5% (to 25.8%) Tau2 (Telecom): +10.1% (to 62.7%) Running on Ubuntu 22, CUDA 13.0 with a single NVIDIA GeForce RTX 4090. Here is the exact step by step benchmarking process with a massive 28k tokens prompt and the commands I used to squeeze out maximum context without killing my throughput: # 1. The Baseline (Unquantized KV Cache) I started with full GPU offload (-ngl 99) and pushed the context to 40k. llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -ngl 99 -c 40000 -fa on --port 8080 -v VRAM: 23.8 GB (maxed out on card) Throughput: Prefill: 2198.81 t/s | Decode: 35.77 t/s (with 28k tokens prompt) # 2. The CPU Split Trap I tried stretching to 80k context by offloading layers to the CPU (-ngl 52). llama.cpp flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 80000 -ngl 52 -fa on --port 8080 -v Throughput: Prefill: 1212.73 t/s | Decode: 5 t/s (with 28k tokens prompt) # 3. The KV Quantization Breakthrough Instead of spilling layers to the CPU, I kept the model fully on card (-ngl 99) but enabled 8-bit KV cache quantization to free up VRAM. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 100000 --cache-type-k q8_0 --cache-type-v q8_0 -ngl 99 --port 8080 -v VRAM: 23.9 GB Throughput: Prefill: 2139.68 t/s | Decode: 32 t/s (with 28k tokens prompt) Result: 100k tokens of context on a single GPU with practically zero speed loss (and minimal intelligence loss). # 4. The Limit Test (Q4 KV Cache) To find the absolute breaking point, I dropped the KV cache to 4 bit (q4_0) and set -c 190000. flags: ./build/bin/llama-server -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -c 190000 --cache-type-k q4_0 --cache-type-v q4_0 -ngl 99 --port 8080 -v VRAM: 23.8 GB Throughput: Prefill: 2206.66 t/s | Decode: 33 t/s (with 28k tokens prompt) (Note: Pushing it to 220k required dropping to -ngl 58 again, which immediately penalized decode down to 17 t/s). # The Tradeoff: For Max Reasoning: Keep your KV cache unquantized (f16). You get pristine reasoning but hit a strict 40k context ceiling. For Massive Document Retrieval: If you need to feed the model giant codebases, use --cache-type-k q4_0. Getting 190k context at 33 tokens/second on a consumer desktop with a 31b dense model is a cheat code. If you’re rocking a single 3090 or 4090 and slept on Gemma 4 earlier, this update is your cue to dust off the terminal. Hugging Face links to the Unsloth QAT quants are in the replies below.

Alok

75,205 views • 9 days ago

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon
0:22

a new 8GB VRAM GPU dense Local LLM leader was born yesterday runs on: RTX 4060 / RTX 3070 / RTX 2080. any 8GB card Qwen 3.5 9B (dense) was the go to for 6-8GB VRAM builds. Gemma 4 12B QAT (dense) just changed that. same llama.cpp + cuda 13.2. i7 12700H. 16GB RAM. same -ngl 99 flags. same 48k context. unsloth gemma-4-12b-it-Q4_K_M.gguf → 15 tok/sec @ 48k ctx unsloth gemma-4-12B-it-qat-UD-Q4_K_XL.gguf → 32 tok/sec @ 48k ctx → 26 tok/sec @ 64k ctx 64k context is a big deal. Hermes 3 agent requires 64k minimum to run. you're now getting full hermes compatible context on a budget consumer GPU at 26 tok/sec locally. 2.1x faster on identical hardware. and here's the part that breaks your brain: the QAT-UD-Q4_K_XL is actually SMALLER than the Q4_K_M "XL" why? QAT = Quantization Aware Training Google didn't train the model first and compress it later they trained it to be quantized from day one the weights already know how to survive low precision that's why you get more quality per byte llamacpp flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -cnv -ngl 99 -c 48000 -v fits in 8GB VRAM clean. no API. no cloud. no subscription. and this isn't even the MTP variant yet Gemma-4-E2B QAT runs on 3GB RAM, E4B on 5GB, 12B on 7GB, 26-A4B on 15GB and 31B on 18GB. I have benchmarked the 26b and 31b qat as well on a single RTX 4090, checkout the comments for details. If you have a 6GB or 8GB VRAM GPU, post your numbers. more benchmarks and configs coming soon

Alok

259,993 views • 1 month ago