Le Chonk goes open: Mistral releases the weights of its 1-trillion-parameter Large 4 🔓

Earlier this month Mistral showed Mistral Large 4, nicknamed "Le Chonk", as a public preview. On 29 October they did the part I was hoping for: they released the weights. Anyone can now download a frontier-class, European-trained model from Hugging Face and run it on their own infrastructure. As someone who loves open models, this one made me smile.

What's inside

According to the model card, Large 4 is a natively multimodal Mixture-of-Experts model with about 1 trillion parameters in total, but only 49 billion active per token (52 billion including embeddings and output layers). It has 128 routed experts, of which 4 are picked per token, plus 1 shared expert that always runs. It reads text and images, writes text, handles a context of 1 million tokens and is natively fluent in more than 160 languages, including every official EU language. So yes, Dutch too.

Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe. You can pick a reasoning effort per request (low, medium or high), and the checkpoint ships with a small built-in draft head for speculative decoding, which Mistral measured at up to 1.9x higher throughput.

Three weeks of extra training

What I find really interesting is how much changed since the preview. Mistral says just three more weeks of reinforcement learning pushed the model forward. On the Artificial Analysis Coding Agent Index it now scores 48%, effectively tied with the leading open model GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it jumped 13 points to 41% and used 30% fewer output tokens.

Security is the other big theme. Mistral reports 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching real vulnerabilities. At the same time, the model refuses malicious cyber requests 96.6% of the time on average, the highest Mistral measured among open models. On the B3 agent security benchmark it resisted 97.1% of attacks, up from 93.3% in the preview.

Why I find this exciting

For me, this is about control. A team can now inspect, fine-tune and host a model of this class themselves, which matters a lot for defenders, banks and public institutions that can't send their data just anywhere. I liked seeing ABN AMRO's director of cyber defense quoted in the announcement, calling open and transparent AI ecosystems essential for innovating responsibly.

The ecosystem was ready on day one too: vLLM, SGLang, transformers and llama.cpp, plus fine-tuning with TRL, Unsloth and Axolotl. Hugging Face even tested a 2-bit quant on a single Apple M3 Ultra and got about 15 tokens per second.

Honest notes

This is not a homelab model for most of us. That 2-bit version still needs a Mac with 512 GB of unified memory, and the full model is meant for multi-GPU servers. The licence is a modified MIT: free to use, modify and commercialize, but companies that offer it as a model-as-a-service need a separate licence. Most benchmark numbers come from Mistral itself, and the full technical report is still "coming soon". Mistral is also open about the risks: it saw potential uplift for real attacks in live traffic and is tightening its guardrails.

Sources

A trillion parameters, open weights and trained in Europe: I'm curious what the community will build, distill and fine-tune from this one.

Reacties

Een reactie posten

Populaire posts van deze blog

Meta just shared the full recipe for a 30B agent model you can run at home 🧠

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while. What the model is Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards. The parts I found most interesting Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compu...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

A 35B model on a Mac mini, using just 2.9 GiB of memory 🚀

There's something special about running a capable language model on hardware you own: no API bill, no data leaving the house, just a small box on your desk doing the thinking. A paper that's currently #2 on Hugging Face's trending list takes that idea further than I expected: it runs a 35-billion-parameter model on a Mac mini at about 20 tokens per second while using under 3 GiB of memory. The trick: leave the experts on the SSD The paper is called The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction , and the system is named Edge0 . It builds on a Mixture-of-Experts (MoE) model, Qwen3.6-35B-A3B. In an MoE, each token only uses a few "experts" per layer (here 4 out of 256 in each of 40 layers), so most of the weights sit idle at any given moment. Edge0 takes that literally. The expert weights (quantized to int4) stay on the SSD and are streamed in only when needed. The clever part is a small prerouter : a tiny trained he...
Meer lezen