Mistral AI Releases the Open Weights of Mistral Large 4, Pitched at Coding and Cyber Defense

Mistral AI on October 29, 2026 released the open weights of Mistral Large 4, the 1 trillion parameter model it nicknames "Le Chonk," on Hugging Face. The released version had three more weeks of reinforcement learning than the API preview and is the best open weight model for coding, the company said.

Mistral launched Mistral Large 4 as an API preview on October 6, 2026 and promised the weights by the end of the month, after red teaming with security firms and governments. The release lands in a crowded field of open weight models, where Chinese labs such as Z.ai, DeepSeek and Moonshot AI hold the top spots on most independent rankings.

What did Mistral release on October 29?

Mistral published the Mistral Large 4 weights on October 29, 2026 under the Mistral Ecosystem Open-Weight License, a modified MIT license. It allows free use, modification, fine-tuning and commercial use, but companies that run a model as a service business need a separate commercial license from Mistral, according to the announcement and the model card.

The model card describes a mixture of experts model, a design that runs only a few expert subnetworks per token: about 1 trillion parameters in total, 49 billion active per token (52 billion counting embedding and output layers), 128 routed experts with four active plus one shared expert, and a 1 million token context window. It takes text and images and returns text. It ships in FP8 and NVFP4 versions with day one support in Transformers, vLLM and llama.cpp, according to Hugging Face.

The model now has three reasoning modes, low, medium and high, with high as the default; low replaces the "none" mode of the preview.

How much better is it than the October preview?

By Mistral's own measurements, the released model gains most on agentic coding: its Terminal-Bench 4.0 score rose 13 points, from 28% to 41%, with 30% fewer output tokens. Independent evaluator Artificial Analysis measured a smaller overall change, one point on its Intelligence Index, along with large gains in speed and cost.

Measure	Preview	Released model
Terminal-Bench 4.0 (Mistral)	28%	41%
FinWorkBench, finance spreadsheets (Mistral)	67.4%	73.3%
B3 agent security, attacks resisted (Mistral)	93.3%	97.1%
Malicious cyber requests refused (Mistral)	95.3%	96.6%
Intelligence Index v5 (Artificial Analysis)	25	26
Output speed (Artificial Analysis)	111 tokens per second	161 tokens per second
Cost per index task (Artificial Analysis)	$5.54	$2.06

Hugging Face's launch blog still lists the preview figures of 28.3% on Terminal-Bench 4 and 93.3% on B3, so it is worth checking which version a number describes.

Artificial Analysis released version 5 of its Intelligence Index in late October 2026, so the preview's earlier score of 38 cannot be set against the new 26. The released model wrote 110 million output tokens across the full index, against 150 million for the preview (The Decoder).

Is it really the best open model for coding and cyber defense?

Mistral's claim holds on some tests and not on others. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48 against 49 for Z.ai's GLM-5.3 while using 47% fewer tokens per task, which Mistral calls effectively tied. Mistral's charts show it leading open models on Terminal-Bench 4.0 and on Cognition's private FrontierCode 2.0 test, at 19.7%.

Measure	Mistral Large 4	GLM-5.3	DeepSeek V4.1 Flash 1017	Kimi K3
Terminal-Bench 4.0	41%	40%	31%	21%
Coding Agent Index	48	49	46	41
Cyber Index	51%	36%	57%	41%
Intelligence Index v5	26	32	32	n/r

On cyber defense the reading is weaker. Mistral says the model scores 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, but its own chart puts DeepSeek V4.1 Flash 1017 at 57% and MiMo-V2.6-Pro at 56%, so Mistral Large 4 ranks third among the open models shown. The preview scored 50 on the same index, according to The Decoder.

Overall, Chinese models lead: Mistral's model card calls its Intelligence Index score of 26 the highest among western open weight models, six points behind DeepSeek and GLM-5.3. Mistral's charts compare against Kimi K3 rather than Moonshot AI's newer Kimi K3.1 Pro, whose weights are due by November 4, 2026; Artificial Analysis scores K3.1 Pro at 46 overall, 67% on Terminal-Bench 4.0 and 64% on the Cyber Index (The Decoder).

Against closed models the gap is wide: in Surge's blind comparisons cited by Mistral, graders preferred its code over Anthropic's Claude Opus 5.5 in only 22% of cases, and Opus 5.5 scores 46% on FrontierCode 2.0 (The Decoder).

How does Mistral handle the cyber risk of open weights?

Mistral says the model refuses 96.6% of malicious cyber requests on average, the highest rate it measured among open models, while still doing defensive work such as reproducing and patching flaws. In the same post it concedes that live traffic since the preview showed "a potential for uplift to conduct real attacks," and says it is modifying its guardrails and will keep updating them.

Vetted partners such as Orange Cyberdefense are testing a specialized variant for penetration testing, limited to authorized systems under contract, Mistral says. Pierre Stock, Mistral's VP of Science, framed open weights as a guarantee of access for defenders:

Access that can't be cut off by definition because the model is open-weighted.

Moonshot AI took the opposite route for its next open release: it says it removed cyber related training data from Kimi K3.1 Pro's open weights, which lowered its score on an internal exploit benchmark from 90.5% to 81.0%. Mistral instead keeps the cyber skills in the weights and relies on refusal behavior, which, as Hugging Face notes, researchers can now measure directly instead of relying on the provider's description.

What we don't know yet

• When Mistral will publish the full technical report it says is "coming soon."
• How the model compares with Kimi K3.1 Pro once those weights are out and tested side by side.
• What changes Mistral will make to its cyber guardrails, and whether they apply to the downloadable weights or only to its API.
• Whether smaller models built on Mistral Large 4 will follow; Mistral says it will serve as the base for specialized models but gave no timing.