Mistral Large 4's open weights did not land on the October 27 date reported: the repo still counts down to October 31 and the LICENSE file does not exist
AspireAI community · Mistral AI
Mistral Large 4 - unofficially ML4, officially "le Chonk" - went into public preview on October 6 as a 1 trillion-parameter natively multimodal mixture-of-experts model with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. Mistral's own announcement commits only to weights "by the end of the month", but multiple reports put a firm October 27 date on the download. Today is that date, and the weights are still not downloadable.
A direct fetch of the Hugging Face repository today returns a page still titled "Upcoming release", carrying an October 31, 2026 "Current ETA" whose tooltip warns the release date may still change, and listing a single planned artifact with no model files served. The LICENSE file that would state what the model permits does not exist yet - Forkast recorded an HTTP 404 on that path today - so no statement about who may deploy it is available to make. For comparison, Mistral Large 3 on the same account ships under Apache 2.0 with 277 files. Mistral's blog index still shows the October 6 announcement as its newest model post.
The model itself is a preview API today, priced at $1.36 per million input tokens and $4.18 per million output tokens, fluent in more than 160 languages. Mistral's own reported numbers include 82 percent on one Artificial Analysis Cyber Index vulnerability-reproduction test, which it says is the highest of any model, and 93 percent on Cybench - against closed models Mistral says score near zero on that test because they refuse the task. Elsewhere it reports 61.7 percent on DeepSWE v1.1, 59.9 percent on AutomationBench, and 42 percent on Dense 200 visual grounding against GPT-6 Astra's 41 percent. On safety, it reports resisting 93.3 percent of attacks on Lakera's public B3 benchmark.
The honest limit: those benchmark numbers are the vendor's own, from a preview that Mistral itself says is preliminary and expected to change before the weights ship, and the red-teaming period with vetted cyber partners and state authorities is still running. The one thing that is not in dispute today is that nobody can yet download the model it is attached to.
https://mistral.ai/news/mistral-large-4/
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.