KTransformers:671B DeepSeek-R1 on a Single Machine-286 tokens/s Prefill
Hey Hacker News! We are excited to share the new version of KTransformers, a flexible framework designed for cutting-edge LLM inference optimizations! Leveraging state-of-the-art kernels from llamafile and marlin, KTransformers seamlessly enhances the performance of HuggingFace Transformers, making it possible to operate large 671B MoE models or extremely long 1M context locally with promising speed. KTransformers is a Python-centric framework designed with extensibility at its core. By implementing and injecting an optimized module with a single line of code, users gain access to a…
In plain words
KTransformers is a Python framework for optimizing large language model inference on consumer hardware. It enables users to run massive models like DeepSeek-R1 (671B parameters) and handle extended contexts (up to 1M tokens) locally at competitive speeds through optimized kernels. The framework integrates seamlessly with HuggingFace Transformers and provides multiple interfaces including RESTful APIs compatible with OpenAI and Ollama, plus a web chat interface. Developers can add performance optimizations with minimal code changes while maintaining compatibility with existing tools and plugins.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hey Hacker News! We are excited to share the new version of KTransformers, a flexible framework designed for cutting-edge LLM inference optimizations! Leveraging state-of-the-art kernels from llamafile and marlin, KTransformers seamlessly enhances the performance of HuggingFace Transformers, making it possible to operate large 671B MoE models or extremely long 1M context locally with promising speed. KTransformers is a Python-centric framework designed with extensibility at its core. By implementing and injecting an optimized module with a single line of code, users gain access to a Transformers-compatible interface, RESTful APIs compliant with OpenAI and Ollama, and even a simplified ChatGPT-like web UI. For example, it allows you to integrate with all your familiar frontends, such as the VS Code plugin backed by Tabby. To demonstrate its capability, we present two showcase demos: - GPT-4/o1-level Local VSCode Copilot: It runs the huge 671B DeepSeek-Coder-V2's Q4_K_M variant using just 24GB VRAM and 382G DRAM (with two Xeon CPU) on a local machine, with a promising 286 tokens/s for prompt prefill and 14 tokens/s for generation, up to 3~28x speedup. The detailed tutorial is [here](https://github.com/kvcache-ai/ktransformers/blob/main/doc/en...) - 1M Context Local Inference:Achieves 15 tokens/s with nearly 100% accuracy on the "Needle In a Haystack" test via the InternLM2.5-7B-Chat-1M model, utilizing 24GB VRAM and 150GB DRAM, and is several times faster than llama.cpp. Check it out on GitHub: https://github.com/kvcache-ai/ktransformers
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
Launched alongside, February 2025
the whole month →
Screen Studio 3.0▲1,833Beautiful screen recordings with instant shareable links
Growth · 2025 · screen.studio
- IG
I was at FB/Meta from late 2013 to early 2023, mostly working in the compiler/runtime spaces. I got hit in the spring 2023 layoff wave. I immediately started making games in my newfound free time (a lifelong interest, and I even worked in AA(A?) back ca. ~2000), and in October 2023 I stumbled upon the idea of a roguelike pachinko/plinko game inspired by Luck Be A Landlord. Things snowballed quickly, I started talking to publishers, then worked like crazy through all of 2024, almost the hardest I've ever worked in my career, and launched the game in December 2024. It's sold…
Work · 2025
- IB
i wanted to change the habit of reaching for my phone in the morning and doomscrolling away an hour so i built an app to help me. now i have to literally touch grass before accessing my most distracting apps the app is built in swiftui, uses the screen time apis provided by apple and google vision to recognise grass or not i'd love to get your thoughts on the concept.
Life & fun · 2025 · touchgrass.now