🎉 Congrats to Thinking Machines on TML Inkling—a 1T-parameter... open-weight model supported in vLLM from Day 0. Highlights: • Natively multimodal across text, image, and audio • Up to 1M-token context • New architecture with relative attention, short convolutions, and MoE expert sinks • 8 MTP heads for speculative decoding vLLM supports both NVFP4 and BF16 checkpoints, optimized for NVIDIA Blackwell and Hopper, reaching up to 380 tok/s/user on 4× GB200 with MTP. Huge thanks to the Thinking Machines Lab team for the close collaboration 🙏 Read about implementation below 👇show more

🎉 Congrats to MiniMax (official) on releasing the open... weights for MiniMax H3! Day-0 support in vLLM-Omni! One model reads text, images, video, and audio as a single context and returns video with native stereo audio. Text-to-video, first/last-frame, and multi-reference generation, 4 to 15 seconds at up to 2K and 24 FPS. The MP4 comes back with H.264 video and a synchronized stereo track already muxed in. It serves over the OpenAI-compatible /v1/videos endpoint, synchronous or async with job polling. 🔊 The video ⬇️ was made with H3, served on vLLM-Omni.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

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

If you thought the Gemma 4 31B (dense) model... was fast, sit down. I just benched the updated Gemma 4 26B A4B MoE on a single RTX 4090 (24 GB VRAM) 9,200 t/s prefill. 160 t/s decode. 250,000 context window. All on a single consumer RTX 4090. The numbers are completely unhinged. The 31B is a dense behemoth. But the 26B is a Mixture of Experts (MoE), specifically an Active 4 Billion (A4B). It holds 26B parameters of knowledge but only activates 4B per token. Because its inference memory footprint is so light, I didn’t even need KV cache quantization to hit a quarter million context. Compiled the latest llama.cpp from source on Ubuntu 22 (CUDA 13). Fed it a 28k token prompt, and manually cranked the batch sizes (-b 2048 -ub 2048) to absolutely redline the Tensor Cores. Here is the benchmarking breakdown: # 1. The Baseline (No MTP) Even without speculative decoding, the A4B architecture flies. llama.cpp flags: ./build/bin/llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 250000 -ngl 99 -fa on -b 2048 -ub 2048 --port 8080 -v Context Ceiling: 250,000 tokens (21.5 GB VRAM) Prefill: 9,200 t/s (Absurd) Decode: 124 t/s # 2. The MTP Overdrive Injected the new MTP draft model to enable Speculative Decoding. llama.cpp flags: ./build/bin/llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp --spec-draft-model mtp-gemma-4-26B-A4B-it.gguf --spec-draft-n-max 4 --spec-draft-p-min 0.7 -c 250000 -ngl 99 -fa on -b 2048 -ub 2048 --port 8080 -v Context Ceiling: 250,000 tokens (22.96 GB VRAM) Prefill: 7,054 t/s (MTP draft overhead slightly caps prefill) Decode: 156 t/s # The Agentic Architecture Insight Why does this matter? Because you can now build a killer local agentic loop on a consumer desktop. Use the 31B dense model (from the previous post) as your heavy, deliberate Orchestrator / Verifier / Planner. Pass the actual execution tasks to this 26B MoE. At 160 t/s, this MoE can chew through code generation, tool calling, and massive RAG document retrieval over a 250k context window almost instantly, drastically speeding up your agentic loop. If you own a single RTX 3090 or 4090 and haven't tried this specific stack yet, you need to pull these latest updates and run it. Local inference just leveled up. Hugging Face links to the Unsloth 26B QAT quants and MTP drafters are in the replies. performance graphs also available in the replies.show more

Gemma 4 12B QAT (dense) achieves 1000+ tokens/sec prefill... on 8GB VRAM with 120k context Gemma 4 12B QAT (dense), TurboQuant (Without MTP), RTX 4060 8GB VRAM: Prefill: 1000+ tok/s (42% increase) Decode: 25+ tok/s (25% increase) Context: 120k (150% increase) prefill was 700 tok/sec and decode 20 tok/sec with only 48k context without turbo quant (older test with mtp link in the comments) llama.cpp TurboQuant flags: -m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf -c 120000 --cache-type-k q8_0 --cache-type-v turbo3 -ngl 99 --port 8080 tested with a 27k prompt, 120k context loaded. -ngl 99 here isn't a typo, full 12B dense, every layer on GPU, on an 8GB card. that's the part worth sitting with. The model has vision, audio input, thinking/reasoning and fits your 8GB card. TurboQuant's KV cache savings are what free up the room to do that at 120k context. side by side with yesterday: 26B A4B MoE got 320+ tok/s prefill. this dense 12B is clearing 1000+ rig: RTX 4060 8GB · i7H · 16GB RAM same two flags as yesterday, different model size: --cache-type-k q8_0 --cache-type-v turbo3 thanks to TheTom/llama-cpp-turboquant, TurboQuant fork of llama.cpp by Tom Turney (Tom Turney) to make this work. unsloth's model quant huggingface and the llama.cpp fork github link in the comments Do you prefer a dense or a MoE for your 8GB card?show more

Run Gemma 4 26b MTP on 8 GB VRAM... GPUs at 25+ tokens/second. Flags included! local llm space is moving at terminal velocity. only 3 days ago google released gemma 4 26b a4b qat quants. more efficient than before, ran on 8gb vram at 20 tok/sec. and now just a few hours ago, mainline llama.cpp merged a massive update and we just shattered our own record. decode throughput went 25-40% up on the same 8 GB VRAM setup! Before MTP: 20 tps -> After MTP: 28 tps! llama.cpp just officially merged PR #23398 ("add Gemma4 MTP"), bringing native Multi-Token Prediction (MTP) support to Gemma 4 models. By running speculative drafting on the same 8GB VRAM RTX 4060 setup, my decode throughput on a 64k context instantly leaped to a blistering 25–27 tokens/sec thats 25-30% increase with the same hardware. Here is the architectural catch you need to know: Unlike the Qwen 3.5 and 3.6 series, which bake the MTP heads directly into the base GGUF, the Gemma 4 MTP head is not built in. You must download a separate, specialized MTP drafter GGUF (the assistant model) to act as the speculator. (I've dropped the download link in the replies). copy and try the exact flags: -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.7 --spec-draft-model gemma-4-26b-A4B-it-assistant-Q4_0.gguf -c 64000 -v n-max 4 and p-min 0.7 is also worth checking out. benchmark on your setup and workflow. if you have a single 8 gb vram nvidia rtx 4060, 3060, 3070, 2080, 2070, grab the MTP drafter GGUF link in the comments and try it yourself. Check it out even if you have asmaller or a larger gpu, such as a single rtx 3090, 4090, 3060, 2060. MTP works for all gemma 4 sizes such as gemma 4 12b, gemma 4 31b etc. but remember to grab the correct mtp draft assistant models respectively. what are you benchmarking todayshow 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

Before the week ends, let's acknowledge one of the... most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality: 🧠 LLMs → NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models. → Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week. → StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0. → Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week. → JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0. 🎨 Image gen (the surprise of the week) → Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights. 🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped) → Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA. → RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0. → Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos. → NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B. 👁️ Vision & VLMs → PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0. → Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0. 🎬 Video, 3D & World Models → NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI. → JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3. → ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).show more

vllm-exl3 v0.3.0 is LIVE with custom native CUDA kernels... for 2-bit EXL3 on NVIDIA DGX Spark GB10. GLM-5.3-Flash-EXL3-K2 jumped from 16.9 → 24.6 tok/s average single-stream decode, a +45.6% gain. Coding hit 27.6 tok/s, +85.6%. 🚀 The previous ExLlamaV3-backed path inside vLLM was leaving a lot of GB10 bandwidth on the table. So I rewrote the hot path specifically for EXL3 on Blackwell sm_121: → in-register Trellis dequantization → native fused MoE decode → power-of-two chunked prefill GEMM → parallel NVMe pre-warm Then I tested it side-by-side on physical DGX Spark hardware using my GLM-5.3-Flash-EXL3-K2 pack and live vLLM HTTP streaming. 🚀 𝗗𝗘𝗖𝗢𝗗𝗘 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧 Single-stream C1: Coding 14.9 → 27.6 tok/s +85.6% Prose 13.7 → 24.6 tok/s +79.3% Reasoning 18.9 → 25.1 tok/s +32.7% Summary 17.1 → 25.6 tok/s +50.0% Format 16.3 → 24.0 tok/s +47.7% Average: 16.9 → 24.6 tok/s 𝗡𝗘𝗧 𝗚𝗔𝗜𝗡: +45.6% ⏱️ 𝗙𝗜𝗥𝗦𝗧-𝗧𝗢𝗞𝗘𝗡 𝗥𝗘𝗦𝗣𝗢𝗡𝗦𝗜𝗩𝗘𝗡𝗘𝗦𝗦 Coding TTFT: 2,344 ms → 859 ms That is a 63.3% reduction, or about 2.7× faster to first token. Follow-up turn with prefix cache hit: 5,608 ms → 3,588 ms 1.56× faster. ⚡ 𝗪𝗛𝗔𝗧 𝗖𝗛𝗔𝗡𝗚𝗘𝗗 𝗢𝗡 𝗧𝗛𝗘 𝗚𝗣𝗨 40 routed-MoE layers: 19.9 ms → 11.5 ms per token Per-layer MoE compute: 497 μs → 287.8 μs That removes 8.4 ms of MoE compute from every generated token. Total per-step wall time: 59.2 ms → 40.6 ms -31.4% The key is `p2b_fused_moe`. Instead of expanding EXL3 weights through a traditional intermediate path, the new kernel performs Trellis dequantization in-register while executing the routed expert computation. The weights stay compressed until the GPU actually needs them. 🔥 𝗣𝗥𝗘𝗙𝗜𝗟𝗟 𝗚𝗢𝗧 𝗔 𝗡𝗔𝗧𝗜𝗩𝗘 𝗣𝗔𝗧𝗛 𝗧𝗢𝗢 The new `exl3_gemm` uses power-of-two chunked prefill GEMM. Measured: 7.85 TFLOPS 13.0× faster than the legacy prefill kernel 1,875 tok/s cold prefill sustained across 65K context 💾 𝗧𝗛𝗘 𝗕𝗢𝗢𝗧 𝗣𝗔𝗧𝗛 𝗡𝗘𝗘𝗗𝗘𝗗 𝗪𝗢𝗥𝗞 𝗧𝗢𝗢 Loading a ~91 GiB model is part of the user experience. Standard shard loading is mostly serial. The updated recipe parallelizes NVMe pre-warm across 8 workers so the storage controller gets used properly instead of feeding a ~100 GiB model one shard at a time. That turns boot-time storage into another optimization target instead of something we simply accept. 💡 𝗧𝗪𝗢 𝗦𝗘𝗥𝗩𝗜𝗡𝗚 𝗙𝗟𝗔𝗚𝗦 𝗪𝗢𝗥𝗧𝗛 𝗞𝗡𝗢𝗪𝗜𝗡𝗚 `--long-prefill-token-threshold 1024` Prevents giant prefill chunks from monopolizing step budgets and starving parallel decode sessions. `--enable-prefix-caching` Avoids paying for the same conversational prefix again on follow-up turns. 📦 𝗘𝗩𝗘𝗥𝗬𝗧𝗛𝗜𝗡𝗚 𝗜𝗦 𝗢𝗣𝗘𝗡 vllm-exl3: GLM-5.3-Flash one-Spark recipe: Model: This is why I like working at the kernel level. The model did not change. The quant did not change. The hardware did not change. The execution path did. 16.9 → 24.6 tok/s. 🛠️ vLLM turboderpshow more

As promised, instead of for $15,000 GPUs, we max... optimized a model that can fit comfortably in any GPU with 8GB VRAM running Windows 200 tok/s, *128k context, text/image/audio. Anyone who wants to join the local AI community with 5 year old laptop can run this with recipe below, and this test was done on 3070ti laptop I forgot I had. We pulled every instruct model released since May that fits 8 or 12 GB, and ran the six that mattered through the same 20 tasks: JSON extraction, code, a table, arithmetic, a fake-treaty honesty check, tool calls, language, a persona, a fact planted deep in a long prompt, and three photos. Gemma 4 E4B, Gemma 4 12B, Qwen3.5-9B, Qwen3.5-4B, Ornith-1.5-9B, the Qwen3.8-9B distill. E4B is not the top scorer. Qwen3.5-9B and Ornith beat it on long context and on reading fine print in a photo. It won the pick for the things a newcomer on an old card actually needs: Apache 2.0, text + image + audio in, 3 GB of headroom on an 8 GB card, Weights Google trained to be 4-bit (QAT), and Google's own speculative drafter shipped next to them. Nothing else in the tier has all five, and the drafter alone is worth 2–3×.show more

Meet Stable Audio 3.0, the open-weight model family built... for artistic experimentation. This is our open invitation to experiment with generative audio. We believe the best innovations are still waiting to be built. The 4-1-1 on 3.0: 📣 You own your outputs, and can distribute and commercialize them under the Stability AI Community License (up to $1 million in revenue). 🎵 New and improved capabilities include variable-length generation up to six minutes, and full song composition on portable devices, no GPU required. ✅ Trained on a fully licensed dataset. 🎨 You can customize the models on your own library with support for LoRa training, which we’ve documented for the first time. More on the models 👇show more

Moonshot AI is casually giving developers free daily access... to Kimi K3 😳 no subscription no upfront payment just sign in and start using one of the largest open AI models available what you get for $0: - Kimi K3 with 2.8T parameters - 1M token context window - strong coding and reasoning performance - native vision capabilities - free daily credits that refresh automatically why this is worth checking: > access a frontier model without paying API fees > long context for large codebases and documents > works on web, desktop, mobile, and CLI getting started takes less than 2 minutes: 1. go to 2. create a free account 3. Kimi K3 is available as the default model 4. start chatting or coding with your daily free credits bonus: Moonshot Together lets you invite friends for a chance to earn 3, 7, 15, 30, or even 365 days of Kimi Membership through its rewards program benchmark highlights: > 2.8T parameter MoE model > 1M context window > strong performance across coding, browsing, and reasoning benchmarks important: free credits reset daily, rate limits apply on the free tier, and the open-weight release is expected on July 27 A simple way to try one of the latest frontier AI models without paying for API accessshow more

my 8 GB VRAM gaming laptop is absolutely going... to hate me for this. but I still did it. ran a 31b dense model (Gemma 4 31b Q4) with only 8 GB VRAM last week I ran Gemma 4 26B A4B a mixture of experts model on my RTX 4060 and hit 25–28 tokens/sec using llama.cpp's new MTP support. smooth. snappy. but MoE has a secret: it only activates 4B parameters per token despite having 26B total. that's why it flies. so the real question started haunting me. what if I throw a full, no tricks, every parameter fires on every token, 31B DENSE model at the same machine? # Hardware: GPU: NVIDIA RTX 4060, 8 GB VRAM RAM: 16 GB CPU: Intel Core i7 H Laptop. Gaming. Modest. The model: gemma-4-31B-it-qat-UD-Q4_K_XL.gguf (model's unsloth huggingface link in the comments) This is Google DeepMind's flagship dense model in the Gemma 4 family that can run on single consumer GPU. It packs a hybrid attention architecture, supports up to 256K context natively, and is QAT (Quantization Aware Training) optimized, meaning it retains far more quality than standard post training quants at the same bit depth. This is NOT the MoE. This is 31 BILLION dense parameters, every single one of them loaded. # the flags I used: -m gemma-4-31B-it-qat-UD-Q4_K_XL.gguf -cnv --spec-type draft-mtp --spec-draft-model mtp-gemma-4-31B-it.gguf --spec-draft-n-max 8 --spec-draft-p-min 0.6 -c 6000 -v Multi Token Prediction (MTP) is still active here. Separate draft GGUF required, same as the 26B setup. # Results: → Decode: ~3 tokens/sec → Prefill: ~2 tokens/sec → Context: 6000 tokens → Hardware crying quietly in the corner: yes so is 3 tps actually usable? For real time back and forth chat? Not ideal. You're not having a fluid conversation at 3 tps. but slow ≠ useless. And this is where it gets genuinely interesting. think about how senior devs actually work in a real team. But when something is architectural, deeply complex, or needs serious reasoning? they walk down the hall and escalate to the senior. That's exactly the local AI agent architecture this unlocks: → Fast orchestrator model (Gemma 4 26B MoE at 25+ tps) handles routing, simple queries, tool calls, memory. The junior dev. → Gemma 4 31B dense is the senior, called only when the fast model genuinely hits a wall. Hard multi step reasoning. Complex code generation. Deep architectural decisions. The agentic loop stays fast. Only the hard hops touch the 31B. That's a legitimate production grade local AI architecture on a budget hardware. (requires 2 8gb gpus) other workflows where 3 tps is completely fine: - overnight batch jobs. summarize documents, extract structured data, review code. Fire it off. Sleep. wake up to results. - One shot deep reasoning - Silent code audit loops, you write and test, the 31B reviews diffs and flags issues in the background between your sprints - Any workflow where output quality > output speed A few weeks ago, nobody was running a 30B+ dense model on a single consumer GPU with 8 GB VRAM. At all. Now we're doing it on an Intel i7-H gaming laptop with a NVIDIA RTX 4060, thanks to llama.cpp + QAT quants + MTP speculative drafting. Google DeepMind said the Gemma 4 31B targets "consumer GPUs and workstations." They were not exaggerating. The hardware bar to run serious frontier class models locally keeps dropping. the tools are here. the models are here. you just have to be willing to abuse your laptop a little. what workflows would you actually run on a local 3 tps 31B dense model? genuinely curious. drop it below.show more

Today we’re open-sourcing Stable Audio Open Small, a 341M-parameter... text-to-audio model optimized to run entirely on Arm CPUs. This means 99% of smartphones can now generate music-production samples in seconds, right on-device with no internet required. Built for fast, on-the-go creation, it turns your next quick idea into up to 11 seconds of audio. Generate drum loops, foley, riffs, and textures right where you are. No cords 🔌 just chords 🎹 You can learn more here:show more

We are in an insane run of open-weight drops.... Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B MoE refresh, Terminal-Bench 2.1 jumps 61.8→82.7 over the preview, DeepSWE 7.3→54.4. Closes in on Opus-4.8 on Agents' Last Exam (25.2 vs 25.7). MIT. → Muse-Glimmer-30B, from Meta (they are back!!): their first open agentic model. ~29.6B dense + perception encoder, 131k+ context, built to run fully local, no cloud. Apache 2.0. → Liquid AI LFM2.5-2.6B: 2.69B params, 131k context, 220 tok/s on an M5 Max in under 2.5GB RAM. Competitive with models 4x larger on agentic tasks. → inclusionAI Ling-3.0-flash: 124B total, only 5.1B active, ~12% the size of their old 1T flagship Ring-2.6, matches it on key benchmarks. MIT. → inclusionAI Ling-3.0-tiny: 7.9B total, 1.3B active, 86-90 tok/s on an M4 Pro MacBook at ~8GB peak memory. MIT. → NVIDIA Nemotron-3.5-Lightning-30B-A3B: hybrid Mamba-2+MoE+Attention, up to 1M context, runs on a single H100 or DGX Spark, SWE-bench Verified 52.8. → deepgrove maple-preview: 20B-A1B ternary-weight reasoner, 218 tok/s on a Mac mini M4, 5.3GB checkpoint. MIT. → BigBang-v1 (endless-frontier): fine-tuned from Qwen3.6-35B-A3B via a self-evolving generator/critic synthetic-data loop. Lands aggregate performance between DeepSeek V4 Flash (284B) and V4 Pro (1.6T), at 35B. Apache 2.0. 🎬 Video → MiniMax-H3: 33B dense omni model, native stereo audio, up to 2K/15s. 3.6k+ likes already. → Minimax-H3-Turbo (lightx2v): Apache-2.0 turbo distillation of H3 for fast inference. → Lightricks LTX-2.5: image-to-video update, custom Gemma-4-12B text encoder, a markedly stronger distilled model. 🔊 Voice → NVIDIA NemotronLabs VoiceChat-11B: full-duplex speech-to-speech, ~450ms turn-taking, #2 on open VoiceBench, and the first open full-duplex model with live tool-calling mid-conversation. 🛡️ Safety → Mistral Shieldstral-1.0-3B: 3B multimodal guardrail that takes your safety policy as plain text instead of fixed categories. Beats LlamaGuard-4-12B and ShieldGemma-9B on HarmBench (99.4) and ToxicChat (84.1) at a fraction of the size. Apache 2.0.show more

Qwen 3.8 Flash decisively crushes the M5 Ultra with... this new recipe by vr8vr8 . I think I found the fastest dual DGX Spark recipe for Qwen3.8-Flash-Next. 138 tok/s structured, 133 code, 383 structured aggregate at four streams, and now it talks to Hermes. It feels like a cloud API, running on your desk. This should torch any current M5 run right now. This is part two. Same pair of Sparks, vr8vr8's recipe, his new v4.1, plus a patch of mine that's now an open PR against his repo. WHAT INITIALLY BROKE Through Hermes, the recipe answered in garbled mixed Chinese, then looped on "function86 to90 turn91" until it hit the token cap. Most people would blame the quant. The quant is fine. Three things in the serving stack were stacking up: Hermes doesn't send a temperature for this model. When a request omits it, vLLM 0.30 applies the checkpoint's own generation config: temperature 1.0, top_p 0.95, top_k 20, sampling on. The documented opt-out flag still leaves the server default at 1.0. Every Hermes turn was sampling hot. Hermes sends reasoning as a top-level reasoning object. vLLM 0.30 only maps a field called reasoning_effort onto thinking. So "thinking off" never arrived. Thinking is on by default in the chat template, and with no effort set it defaults to xhigh. A request for no thinking became the longest think the model can do, at temperature 1.0. This is why benchmarks alone do not tell you if a recipe is ready. A real harness that leaves temperature out and names the reasoning field differently gets a different model. Mia's older image never applied the generation config and drafts with argmax, so omitted temperature never sampled hot there. This recipe samples the draft. THE PATCH Two changes to vLLM's chat protocol, applied on both ranks by The Hermes reasoning object now sets thinking. none turns it off. low, medium and xhigh pass through to the template. An explicit enable_thinking still wins. If a request omits temperature, the chat path uses 0. An explicit temperature still passes through. Receipts: three table requests with no temperature and thinking off came back in English. A medium request closed its think and answered in English. The PR is open, not merged. Until it lands, a stock Hermes custom provider still samples at 1.0. HIS V4.1 vr8vr8 shipped recipe v4.1 in the meantime, same v6 image: FlashInfer GDN prefill. His card claims about +5% prefill with decode unchanged. On my depth ladder, prefill came in +6 to +7% from 128k to 900k. Claim holds. THE SECOND GRID Same pair, same protocols, thinking off, temperature 0. My runners already send temperature 0, so this is v4.1 against part one, not a test of the patch. Prose 600: 81.6 tok/s (part one 78.4) Structured count 200: 138.3 (131.8) Code clamp: 133.2 (126.4) JSON catalog: 118.6 (111.8) That's up 4 to 6% across the board, which is run-to-run band. I'm not crediting FlashInfer for it. At four streams: structured 383.1 aggregate, code 330.5, JSON 290.4, prose 200.5. Single-stream structured costs 0.52 J per token of GPU power across both boxes. Tool-call smoke test clean, finish_reason tool_calls, no XML in content. DEPTH Needle retrieval 18/18, from 8k out to the 900k rung (880k server-measured). Decode after fill holds between 65 and 77 tok/s the whole way, 70.6 at the top. Prefill runs 3,245 tok/s at 128k and still 2,577 at 880k. AGENT TURNS The OMP agent lane now runs a full 2,048-token write_file turn at 35k context, replacing part one's 35-token window: 107.4 tok/s decode. The Hermes research turn, two web_search calls at 35k context: 81.0 tok/s, up from 74.7. STILL NOT LIKE-FOR-LIKE Against Mia's d2f54b7 recipe on this same pair: prose 1.6x, structured 2.0x, code 2.0x, JSON 1.9x, Hermes turn 2.2x. Different quant, K=5 vs K=3 drafting, bf16 KV vs fp8. My own version is also my own customized recipe, but it felt better to just serve a PR to the creator, since I didn't invent anything new. The biggest single lever is his output head: 1.18 GiB down to 0.33 GiB, read about 5.4 times every decode step, worth +30% engine steps on its own. SUMMARY Every clean lane faster than the mainstream recipe on this rig: checked. Retrieval 18/18 to 900k: checked. Decode holds to 900k: checked. Agent lane measured over a full 2,048-token window: checked. Hermes out of the box: PR open, work in progress. The speed was always the hard part, and vr8vr8 had it in part one. Sources , Benchmarks and Recipe in Reply 👇show more

Big moment for text-to-speech. Qwen just open-sourced a text-to-speech... model that lets you clone voices, design new ones, and control speech using natural language. Let me explain what I mean: You can literally tell it "speak in a cheerful tone with slight nervousness," and it actually does that. No complex audio engineering needed. What makes this special: - 3-second voice cloning - Covers 10 languages: English, German, French, and more - Latency as low as 97ms for real-time applications - Supports both streaming and non-streaming generation The model comes in two sizes (0.6B and 1.7B parameters), so you can pick based on your hardware and quality needs. Three modes to work with: 1. Custom Voice: Use pre-built premium voices with instruction-based style control 2. Voice Design: Describe the voice you want in plain English (or Chinese), and the model creates it 3. Voice Clone: Provide a 3-second reference audio and clone that voice The best part? It integrates with vLLM for production deployment and has a simple Python package you can pip install. I've shared a link to the GitHub repo in the next tweet.show more

Run Gemma 4 26B MoE on 8GB VRAM with... 250k context at 20+ tokens/sec If you own any 8GB VRAM graphics card, stop what you are doing. Local AI just had its absolute "Holy Shit" moment for budget hardware. Yesterday, I benchmarked Unsloth Gemma 4 12B Q4_K_XL on an 8GB card. The community went wild but immediately demanded more: "Can we run a 25B+ model on budget GPUs?" Today, I’m delivering exactly that. I am running a massive 26B parameter Mixture of Experts (MoE) model locally on a standard 8GB VRAM setup with 250k full native context!. If you own an RTX 3060, 3070, 4060, or any budget GPU with 8GB of VRAM, the local AI paradigm has completely changed. The performance metrics are astonishing: - 20 tokens/sec flat decode throughput. - Stable, flat decode speed even with massive prompts. - I threw a 60k token prompt at it, and it still clocked in at 20 TPS without dropping a single frame. # What about prefill? Yes, Time To First Token (TTFT) is slightly high when swallowing massive contexts. But with a solid 200 tokens/sec prefill speed, the wait is barely noticeable and highly usable. And this is running completely without Multi Token Prediction (MTP) active. How is this possible? It’s the magic of Google's new QAT (Quantization Aware Training) quants for Gemma 4. The model weight file (unsloth gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf) is only 13.2 GB, making it the ultimate local powerhouse. # The Test Setup: CPU: Intel Core i7 RAM: 16GB System RAM GPU: NVIDIA GeForce RTX 4060 Laptop GPU (8GB VRAM) # The Secret Sauce (The -cmoe Flag) To make this work properly on any 8GB card, you must use the -cmoe (CPU MoE) flag in llama.cpp. This flag isolates the heavy MoE expert weights directly to system memory (CPU/RAM) while letting your GPU focus strictly on the Attention layers and the KV Cache. It prevents VRAM spillage and holds the throughput rock solid. # The flags: -m "gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf" -cmoe -c 248000 -v Once running, just open the UI on localhost and toggle the new reasoning lightbulb icon in the text input box to watch the model perform multi step thinking. Are you still running smaller models, or are you ready to scale up your budget local setups? Let's discuss in the repliesshow 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