Mistral Large 4 Review: Trillion-Parameter Open Weights, Strong in Coding and Security, Mid-Pack Overall

The architecture sets the bar for running it: 1 trillion parameters does not mean every computation uses all of them. Per token, the router picks 4 of 128 experts plus 1 shared expert, for about 49 billion active parameters, or 52 billion counting embeddings and output layers. The text decoder has 72 layers, the first dense and the other 71 mixture-of-experts; attention combines multi-head latent attention with a sparse-attention indexer that selects up to 2,048 earlier positions per query; images pass through a 1.6-billion-parameter Pixtral vision encoder.

1. Publisher and dates: Mistral AI; public preview Oct 6, 2026; open weights released Oct 29, 2026
2. Model: Mistral Large 4, nicknamed Le Chonk, a natively multimodal mixture-of-experts model; text and image input, text output
3. Scale and context: about 1 trillion total parameters, about 49 billion active per token (52 billion counting embeddings and output layers), 1M-token context window
4. Reasoning effort: low, medium and high settings, high by default, and reasoning cannot be switched off
5. Channels: Hugging Face repository mistralai/Mistral-Large-4-1T-A52B in official FP8 and NVFP4 variants, plus community GGUF quantizations; API via Mistral Studio, the Mistral API, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry
6. Price: API at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens; preview pricing was $1.36 input and $4.18 output
7. Licence: Mistral Ecosystem Open-Weight License, an MIT derivative permitting use, modification, fine-tuning and commercialization; businesses offering access as a model-as-a-service must first obtain a separate commercial licence
8. Hardware bar: the official FP8 inference recipe uses 8 NVIDIA H200 GPUs, serving up to 131K tokens of context and 16 concurrent requests; the NVFP4 build needs Blackwell chips, with the official vLLM example spread across 8 cards; a 2-bit community quantization runs on a single Apple Silicon Ultra with 512GB of unified memory

The release figures come from Mistral's Oct 29 release announcement, cross-checked across several reports. Terminal-Bench 4.0 rose from 28.3% at preview to 41% at release, up 13 percentage points, with output tokens down 30% at the same time; Cognition's private FrontierCode 2.0 benchmark reads 19.7%, above DeepSeek V4.1 Flash at 19.3%, GLM-5.3 at 17.8% and Kimi K3 at 17.3%; the Artificial Analysis Coding Agent Index reads 48%, level with GLM-5.3, using 47% fewer tokens per task on average.

Security and knowledge work form the other set of release figures: the Artificial Analysis Cyber Index reads 51%, attack resistance on Lakera's B3 agent-security benchmark reads 97.1%, and the refusal rate on malicious cybersecurity prompts reads 96.6%; FinWorkBench finance spreadsheet tasks read 73.3% and BrowseComp browsing retrieval reads 84.5%. All of these are vendor-reported figures, not yet independently reproduced; Cognition's FrontierCode 2.0 is a private benchmark that outside parties cannot rerun directly.

Preview-checkpoint figures, repeated in the Hugging Face Oct 29 launch post, must be read separately from the release figures: DeepSWE v1.1 reads 61.7%, SWE-Atlas-QnA 59.4%, Terminal-Bench 4 at preview 28.3%, the Coding Agent Index at preview 49.8%, AutomationBench 59.9% and AA-Briefcase 1,393 Elo; on security single tasks, the reproduce-then-patch task reads 82%, Cybench reads 93% across 40 challenges, Lakera B3 at preview reads 93.3%, and visual grounding on Dense 200 reads 42%, above GPT-6 Astra's 41%.

Independent boards give external figures to set against the vendor numbers; their methods differ from the vendor's own runs, so their scores are not folded into a vendor average. Artificial Analysis' Intelligence Index v5 reads 26, fifth among open-weight models and the highest among Western open-weight models; the same board's Nov 1 page lists the Cyber Index at 51, matching the vendor's release figure. LMArena board data dated Oct 28, 2026 reads 1,430 Elo for text and 1,478 Elo for coding. Vals AI's independent Terminal-Bench 4.0 run reads 27.78%, below the vendor's 41%, and the different evaluation harness is the direct cause of the gap.

The conversion is stated openly: one representative vendor-reported score per dimension, divided by 10 into a 10-point scale; the total is the mean of the five dimension scores. The converted scores only restate benchmark figures, with no editorial judgement added; independent board scores are listed separately and excluded from the total.

1. Agentic coding 4.8: Coding Agent Index 48 converted, from Mistral's published release figure
2. Terminal tasks 4.1: Terminal-Bench 4.0 41 converted, from Mistral's published release figure
3. Cybersecurity overall 5.1: Cyber Index 51 converted, from Mistral's published release figure
4. Vulnerability reproduction and patching 8.2: CyberGym-E2E 82 converted, from Mistral's published figure
5. Browsing retrieval 8.5: BrowseComp 84.5 converted, from Mistral's published release figure
6. Total 6.1: mean of the five converted dimension scores (4.8+4.1+5.1+8.2+8.5)/5
7. Independent boards, listed separately: Artificial Analysis Intelligence Index v5 26, fifth among open weights; LMArena text 1,430 Elo and coding 1,478 Elo (Oct 28, 2026); Vals AI Terminal-Bench 4.0 27.78%, below the vendor's 41% on a different harness; none counts toward the total

On function and price, the following teams can already obtain the weights through public channels or connect via the API: security teams that need to reproduce and patch vulnerabilities, run penetration tests and red-team exercises inside their own systems — Mistral also offers a dedicated cybersecurity variant preview to enterprise design partners, with access subject to strict review and a signed agreement and testing limited to their own or explicitly authorized systems; teams with hardware at the scale of 8 H200 cards or 4 B200/B300 cards that need to run a trillion-parameter model on their own hardware; and teams using the API on a pay-per-use basis at $0.70 per million input tokens and $2.10 per million output tokens.

In the other direction, businesses offering access to the model as a model-as-a-service must first obtain a separate commercial licence; teams needing the full 1M context window in deployment should note that the official 8-card H200 inference recipe serves up to 131K tokens of context, so the deployment recipe cap is not the same as the model window, and re-testing on the target configuration comes before deployment; teams selecting on the terminal-task score alone face a harness gap between Vals AI's independent 27.78% and the vendor's 41%, so re-testing on the target task is a necessary step before deployment.