Georgi Gerganov

RT Xuan-Son Nguyen Someone on Reddit made a searching app that runs on EmbeddingGemma 2 🤯 Much better than MacOS's default Spotlight 🔥

RT Xuan-Son Nguyen Replay of my talk at @dotConferences is out! In this talk, I went deep into speculative decoding and dflash/dspark, explaining how it helps improve the token generation speed, and showing how easy it is to use on llama.cpp Thanks the dotConferences team for organizing this!

RT merve if you want to get every last bit of performance out of your local setup, you need to know a bit about inference 🥵 but we've got you covered, shipping conceptual guides 🔥 here's the first one about prefill, decode, KV cache what you should optimize for 🙌🏻

RT QVAC Another PR of ours just merged into llama.cpp: the fast Metal kernels from Monday now cover all 26 weight formats the Mac GPU runs. They made speculative decoding faster on Mac for 10 formats, and this PR adds the other 16, including BF16, MXFP4, Q2_K, Q3_K and the IQ quants. On an M3 Ultra, matrix multiplies run up to 4.4x faster when a model checks several tokens at once. Thanks @ggerganov for asking for it and merging it within a day. http://github.com/ggml-org/llama.cpp/pull/30065

RT 🍉 Abubakar Abid The Windows launch event today has mentioned "local AI" 41 times (so far) 🤯

llama.cpp can distribute inference on heterogeneous devices through the ggml RPC backend It's an advanced setting but I think with time we'll make it more accessible to regular users.

RT merve releasing my local AI slide deck covers from prefill vs decode, MoE vs dense, VRAM vs unified memory, quantization to speculative decoding, everything with llama.cpp feel free to reuse with attribution

Day-0 support for EmbeddingGemma 2 in llama.cpp https://huggingface.co/ggml-org/embeddinggemma-2-GGUF

If you are using Qwen3.8-27B + MTP, make sure to upgrade to DFlash for extra speed: llama serve -hf ggml-org/Qwen3.8-27B-GGUF --spec-type draft-dflash --spec-draft-n-max 7 Requires the latest llama.cpp v0.6.0

RT Victor M decision models + local ai = ❤️ tried Laya (421M) on llama.cpp in a little apple-picking benchmark, it's almost as good as Qwen 4B at 12x the speed.

RT Luis Catacora You can now run Clef via llama.cpp! llama serve -hf ggml-org/Clef-GGUF Ollama vs llama.cpp speed test on my M5 Max MBP:

The new v0.6.0 release packs a lot of good stuff: - Clef support (text + vision) - High-quality support for Qwen3.8-Flash-Next - Massive performance improvement with Metal - New `llama_batch_ext` API Also the website got a nice refresh: https://llama.app

RT Marco Chiappetta We can't thank llama.cpp and @ggerganov enough for being among the pioneers of local AI. But we sure as hell can contribute back! Upstream contributions will now become a priority in @qvac.

RT QVAC One of our PR just got merged into llama.cpp: new Metal kernels for speculative decoding on Apple Silicon. Before this PR, speculative decoding on a Mac was slower than plain decoding. On an M3 Ultra it now runs up to 3.4x faster than plain decoding, 110 tok/s against 32.1. Thanks to @ggerganov for reviewing and refining it. http://github.com/ggml-org/llama.cpp/pull/29869

Decision models in llama.cpp are now available The `/v1/systemone` endpoint is available in the latest llama builds. Use it to do Jev-style inference locally, efficiently and privately. Multiple open models are supported with more to come. https://huggingface.co/blog/ggml-org/decision-models-in-llamacpp

RT clem 🤗 getting acquired by @nvidia = hugging face can now hire people we couldn't as a small startup and give them a decade to make open-source AI win! if you're one of them, my dms are open

RT merve when I write docs you will not know but there will be signs

Run GGUF models directly with transformers. This work brings ggml's Metal kernels to the transformers ecosystem, increasing compatibility and performance. More info below

RT Xuan-Son Nguyen My talk about llama.cpp speculative decoding (MTP, dflash, dspark) at dotAI 🦙🦙 replay available soon!

RT Google Gemma You don't need to be a command-line expert to run local AI. http://llama.app wraps the heavy lifting of llama.cpp into a clean UI, making it simple to deploy and manage models like Gemma 4. Instead of dealing with config files, you get one-click model downloads, clear memory estimates, and no coding required! Watch the video to see Gemma 4 in action as it parses tables from receipts, streams reasoning logs, and connects to MCP for web search.

RT Alejandro AO 🤗 Just made a quick tutorial on running local models in Pi with llama.cpp step by step 0:00 Why local models in Pi 0:59 Install llama.cpp 1:49 Choose and load a model Written version in 🧵

Hugging Face has been acquired by NVIDIA It is quite exciting to be a part of this journey! NVIDIA has been an active supporter of the llama.cpp project. For more than a year now, their engineers have actively contributed to the codebase, collaborated with the community and provisioned hardware for development and testing purposes. The local AI ecosystem has largely benefited from our joint efforts. Going forward, llama.cpp/ggml will stick to its founding principles. One of the most important qualities of the project is to be hardware-agnostic. Therefore the development and support of all backends will continue to be done as usual - driven and shaped by the community. Open to everyone who is willing to participate. The existence of such an independent software platform is crucial for the rapid adoption of AI locally and for bringing it closer to the user. Now, with such a significant partner as NVIDIA supporting us, in addition to the wider Hugging Face team, I believe that our long-term goals will be easier to achieve. We will remain focused on building one of the most exciting projects based on the most exciting technology of our lifetime and making it truly available to everyone in the most accessible and efficient way possible.

RT Julien Chaumond Super happy to officially announce that we are a̶c̶q̶u̶i̶r̶i̶n̶g̶ joining forces with @nvidia 🔥 Here is a more personal take: AI is at an inflection point. Open source AI can become less relevant in the coming years if the big closed labs run away with it, OR it can become the foundational fabric of the next phase of human civilization. Those are vastly different outcomes, and we need the critical mass to ensure we give our collective best shot to the second outcome. Given @JensenHuang's stance on open source AI and how he stepped up to defend it when it was under threat earlier in the summer, NVIDIA was the only partner we truly considered. HF will remain an independently run, neutral platform. This gives fuel to our long-term vision and mission of unlocking the community's progress to ensure that AI, which is the greatest breakthrough of our lifetime, is accessible to as many people as possible.

RT Xenova Introducing Fleet: GPU benchmarking entirely in your browser. Run WebGPU compute kernels drawn from real AI workloads directly on your hardware and earn a personalized card built for your device. Rate your GPU. Join the Fleet. https://webgpu-kernels-fleet.hf.space/

The Llama app for Mac now comes with a simple request builder for llama.cpp's REST API

RT merve llama.cpp docs now have a new home, shipped with ❤️ upcoming: speculative decoding in detail, quantization (k-quant, i-quant), coding agents what do you want to see next? 🤗 📖 http://llama.app/docs @llama_cpp

let that sink in

I find the "inception" pattern to be very useful in many agentic use cases. You can force the model to take an action when it thinks for too long by injecting a thought after a specified reasoning budget. Helps dealing with underspecified tasks which make the model reason for way too long.

RT Georgi Gerganov simple: llama serve -hf ggml-org/Qwen3.8-27B-GGUF --spec-type draft-mtp

RT Xuan-Son Nguyen We are happy to announce that Muse Glimmer is day-0 supported on llama.cpp. Meta also provides an official GGUF quant:

RT Mark Zuckerberg Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

RT merve if you have a mac/book and don't use llama-macos you're ngmi this beautiful piece of widget > gets your mac + your desired context window to recommend models > kicks off llama server + webui > easy switch between open models $ brew install --cask llama-app

RT Pi Hi kids, gramps here, taking over for the intern. Just released pi 0.81.0 which features first class integration with @ggerganov wonderful llama.cpp server. Professional video demo recorded in an echo-y Austrian office below. Shout out to our friends at @huggingface.

Run upgraded Gemma 4 with llama.cpp (QAT + MTP)

RT clem 🤗 We just crossed $100M annual run-rate. I know many AI companies are capturing much more $$$ these days, but still proud of the milestone! Maximizing short-term revenue has never been our priority. In fact, we're proud to manage to store and serve hundreds of petabytes of models and datasets while keeping HF free and open-source for 97% of our users. As a platform, we’re happy to hopefully create orders of magnitude more value for the community than what we capture. To me, that’s the very definition of a platform. And it has helped us build one of the most loved platform in tech, with network effects, a defensible position and a sustainable business which is quite unique in AI. Many many thanks to all the community members for building with us, we wouldn't be anywhere without you! Can’t wait for what’s next, especially as more companies start to see the value of open and local AI! Next milestone $1B?

RT Julien Chaumond Llama.cpp has a new branding + official website. Run local models today! Now more than ever, open source must win. 🙏 By @alekgrygier and @ggerganov at ggml/hf

RT Alek Grygier Check out our new logo and website on http://llama.app ;)

Some cool ggml-based work by @mudler_it recently - make sure to check it out.

RT Google Gemma Building super fast experiences with Gemma just got easier. Gemma 4 MTP is now officially merged into llama.cpp. Developers can now pair MTP with Gemma 4 QAT for a fast, lightweight setup.

Highlighting recent advances in multi-GPU and tensor parallel support in llama.cpp Over the last few months llama.cpp maintainers and engineers from NVIDIA collaborated to improve the multi-GPU performance in ggml. This resulted in significant performance gains on RTX systems and laid the groundwork for hardware-agnostic tensor parallelism in ggml. For more information on this and other advancements in the low-level inference engine of llama.cpp, check the technical blog by @NVIDIARTXSpark below

Strong signal for local AI on this year's Computex. Big players like NVIDIA and Microsoft are embracing and discussing local AI workloads. Dedicated consumer hardware and models are on the way.