Mistral releases open weights for 1-trillion-parameter Mistral Large 4
For the head of model risk at a regional bank
Mistral AI released the weights of its 1-trillion-parameter model Mistral Large 4 ("Le Chonk") on Oct. 29, according to the company's blog post and its Hugging Face model page.123
Mistral said it released the model in public preview earlier this month, then spent about three weeks on more reinforcement learning and customer feedback before publishing the weights.13 The company says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters and pitches it as a base for sovereign, self-hosted AI.12
On the Artificial Analysis Cyber Index, Mistral reported a 51% score for the model; its own chart lists other models at 56% and 57%, and the blog post does not name which models those are beyond the chart labels.1
Mistral said the model's average refusal rate on malicious cyber requests is 96.6%, the highest it measures among open models, based on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm.1
The Hugging Face page says the weights come under the Mistral Ecosystem Open-Weight License, a modified MIT license for commercial and non-commercial use, and that running a Model-as-a-Service business requires a separate commercial license with Mistral.2
Mistral said the model scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, using 47% fewer tokens per task; on the Artificial Analysis Intelligence Index v5 it scores 26, which Mistral says is the highest among western open-weight models, with leading Chinese open-weight models scoring higher.12
Self-hosting the weights lets a bank keep prompts and code inside its own firewall, which settles data residency in a way a vendor contract cannot. The benchmark figures are Mistral's own, the cyber score trails two models on its own chart, and a bank running the model takes on the operating burden itself, so it should test on its own workloads first.
Mistral said it plans to release additional models from this new generation in the coming months, per SiliconANGLE's report of the company's statement.3
Our take. A frontier-grade coding and cyber model can now sit on your own servers, where no vendor sees your data. No vendor validates, patches or answers for it either. The file is yours, and so is the risk.
Action. Model risk: Add self-hosted open-weight models to the model inventory, with a named internal owner for validation, version updates and security patches.
What our records say
NVIDIA sells GPU compute and a licensed AI Enterprise and NeMo software stack that banks run in their own data centres or rent in the cloud, and PNC says owning GPU compute is meant to reduce reliance on external, usage-based pricing.1
From our records
- NVIDIAvendor · banks using it (6) BNY, Amex, Capital One · alternatives SambaNova
Sources
- Primary sourceMistral Large 4 (Le Chonk) is now open weight | Mistral Mistral
- Primary sourcemistralai/Mistral-Large-4-1T-A52B · Hugging Face Huggingface
- Also reported byMistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model Siliconangle