Hugging Face

So excited to speak at #PyTorchCon Europe next week in Paris. I'll be sharing bridging the hardware gap with code agents. tl;dr: everyone is using agents but how do we scale this to your hardware. Wednesday April 8, 2026 15:25 - 15:50 Master Stage

future is local 🔥 Google DeepMind just released Gemma 4: local frontier in many sizes, all modalities with free license 🤯 we ship Gemma 4 in transformers, llama.cpp, transformers.js and more for your convenience 🫡 plug-and-play with your agents 🙌🏻 read our blog ➡️ https://lnkd.in/erT5DefA

BREAKING NEWS: Google just re-entered the game 🔥🔥 They want to take the crown 👑 back from Chinese open source AI. And... Gemma 4 is FINALLY Apache 2.0 aka real-open-source-licensed. From what I've seen it's going to be a pretty significant model. But give it a try yourself today: brew upgrade llama.cpp # you might need to install from source until build 8637 is in your package manager later today: brew install llama.cpp --HEAD 🔴 My personal recommendation: if you have at least 24GB of RAM or VRAM, run the (very good) 26B MOE: llama-server -hf ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_K_M if you have 16GB of RAM or VRAM, run the dense E4B: llama-server -hf ggml-org/gemma-4-E4B-it-GGUF:Q8_0

So happy to see Google release Gemma 4 today in apache 2.0 that gives you frontier capabilities locally. Local (aka on your device) is the future of AI because it’s free, safer & faster than APIs or clouds. It’s also the best way to mitigate current and future compute shortages and distribute control & power! You can use Gemma 4 right away in all your favorite open agent platforms like openclaw, opencode, pi, Hermes by asking it to change your model to local gemma 4 with Llama server. Local AI is having its moment and we’re here for it! Thank you Google Demis Hassabis Sundar Pichai Omar Sanseviero 💎💎💎💎

llama.cpp just hit 100k stars on GitHub 🔥 Here's a bold prediction: within 18 months, 90% of all AI agents will be running locally on your own machine. Here's why local agents are inevitable:  → Models keep getting more efficient. MoE architectures like Qwen3.5-35B-A3B give you frontier-level quality while only activating 3B params at inference.  → Quantization keeps getting better. The work from Unsloth AI among others is outstanding.  → Apple Silicon and MLX made local inference fast and accessible. 64, then 128GB of unified memory will go a long way.  → Agents don't just need 200 tok/s. They also need reliable tool use, good reasoning, and long context — all things local models are getting great at. Your data never leaves your machine. No API calls, no rate limits, no vendor lock-in. 2026 is the year of Local Agents.

Introducing the first discrete diffusion pipeline for text in Diffusers -- LLaDA2 by InclusionAI 🔥 It follows an MoE architecture with 16B total params. It is definitely not SOTA across the board, but it hopefully flips that soon. Diffusion is being explored for a bunch of OCR use cases as well, and I hope this integration will pave the way forward for many more dLLMs to come. Check out the links below to know more.

A small 🎁 to anyone training models this weekend, we just released Trackio 0.20.1 along with an end-to-end tutorial showcasing the fresh new UI! https://lnkd.in/gcaZ5bbi

Many AI workloads need somewhere to dump data that isn't a final dataset yet i.e. intermediate results, batch outputs, collected records that need processing before they're ready to publish.... That's exactly what Storage Buckets are for. Hugging Face just shipped them: mutable, non-versioned object storage on the Hub, powered by Xet. I immediately had a use case. I maintain a pipeline that collects README cards from datasets and models across the Hub. The old setup: - GitHub Actions every few hours - Download the entire existing dataset (multiple GB) - Merge in new cards, deduplicate - Re-upload everything Ev. Every run moved gigabytes of unchanged data,d it was quite slow and brittle because of storage limits in GitHub actions. Storage Buckets as a "working layer" made this much simpler: - Fetch jobs just append JSONL batches to a bucket — no need to read existing data - A daily compile reads from the bucket, deduplicates with Polars, publishes to the Hub - The whole compile takes about a minute for 400k+ records Because Buckets are backed by Xet, you can be lazy about deduplication at the storage level since Xet handles chunk-level dedup for you, so overlapping writes across runs don't balloon storage or transfer costs. You just write and let Xet sort it out. Each stage only writes forward. Fetch never reads the bucket. Compile never modifies the bucket. The published dataset can be regenerated from the bucket at any time. The whole thing runs on HF Scheduled Jobs — UV scripts with inline dependencies, no Docker, no CI config 🤗 Wrote up the pattern here: https://lnkd.in/emxEKrMi

Any Gradio Space is now callable from a terminal one-liner. The Gradio CLI now has predict and info commands.👇 This matters because coding agents (Claude Code, Cursor, Codex) can now call any of the 400k+ Gradio Spaces on Hugging Face without writing a Python script first 🤯