Models · scorecard

Mistral large 4

What the vendor claims, next to what independent boards measured. The two are never mixed in one column.

At a glance

OpenRouter does not list this model, so there is no price or context window here.

What the vendor claims

Quoted word for word from the vendor's own launch post or model card. Claims, not measurements.

Scores the vendor claims, each with its source and the quoted line
Benchmark	Claimed	Source
FrontierCode 2.0 (Cognition, private)The source says: On Cognition's private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge the agent's change, Mistral Large 4 leads with a score of 19.7%.	19.7%	Mistral · Oct 29 From Drop #054 · Thu, Oct 29, 2026 · Evening
Terminal-Bench 4.0The source says: On Terminal-Bench 4.0, Mistral Large 4 improved by 13 percentage points to 41% compared to its public preview, taking the lead among open-weight models while reducing output tokens by 30%.	41%	Mistral · Oct 29 From Drop #054 · Thu, Oct 29, 2026 · Evening
Refusal rate on malicious cyber requestsThe source says: It retains strong performance on Cyber benchmarks, while achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models.	96.6%	Mistral · Oct 29 From Drop #054 · Thu, Oct 29, 2026 · Evening

Independent results

Measured by independent sources (Epoch AI, LMArena, OpenRouter's benchmark runs, τ²-bench, Vectara), not by the vendor. "Listed as" is the source's own name for the model, so you can check our match.

No independent results yet. They appear here as Epoch AI and LMArena add this model.

In the Drops

ModelsOct 30, 2026Following since Oct 29, 2026in 2 Drops

Mistral pitches Le Chonk's open weights at sovereign cyber defense, TechCrunch says

Mistral says Mistral Large 4 performs comparably to China's GLM-5.3 while using 47 percent fewer tokens per task. Weights are on Hugging Face and its API; the HN thread drew 1260 points.

Why it matters: Security teams that can't send code to a closed API now have a trillion-parameter model they can run on-prem, per Mistral's pitch.

TechCrunch · Oct 29 Hacker News · Oct 29

From Drop #055 · Fri, Oct 30, 2026 · Morning

ModelsOct 21, 2026

Mistral, NVIDIA plan engineering AI benchmarks; Large 4 weights planned for late October

Mistral says the pair will build public benchmarks and open datasets for physics models and for agents in CAE and EDA work. Mistral Large 4 has about 1 trillion parameters, 49 billion active, trained on 3,800 NVIDIA GPUs.

Why it matters: Shared benchmarks would let aerospace, auto and chip teams compare engineering agents on real tool use, with execution time and compute cost measured.

Mistral · Oct 21

From Drop #038 · Wed, Oct 21, 2026 · Evening