Mistral Large 4 Weights: What Enterprise Buyers Still Need to Verify

Mistral released the Large 4 weights on October 29, 2026, after an October 6 preview, giving companies that need open-weight access a coding model to evaluate. ² The two sources report different Terminal-Bench results, and neither reports cost or performance on a buyer’s own workload. ¹²

Hugging Face describes the model as having about one trillion total parameters, 49 billion active parameters, and a one-million-token context window. It accepts text and images and generates text; Hugging Face calls it multilingual and multimodal. ¹ Those details can inform an initial evaluation, but they do not establish which variant a buyer can legally run or its quality and operating cost on a particular workload.

What changed in the Mistral Large 4 weights release

Hugging Face hosts the Large 4 checkpoints in a dedicated collection. The source lists FP8 and NVFP4 variants, as well as community GGUF versions created by Unsloth and the llama.cpp team. ¹ These are distinct releases; the source provides no comparative quality or resource measurements for them. Hugging Face lists Transformers, vLLM, and llama.cpp for inference, TRL for fine-tuning, and its own inference endpoints. ¹

Hugging Face describes a mixture-of-experts architecture with sparse attention and a Pixtral vision encoder: a 72-layer text decoder paired with a 1.6-billion-parameter vision module. ¹ These details can help frame a technical evaluation, but the source does not specify the serving setup required to meet a particular latency or throughput target. ¹

Mistral Large 4 coding benchmarks don’t match

Mistral positions Large 4 for coding and cybersecurity work. ¹ The sources report different headline benchmark figures, without explaining the discrepancy. ¹²

Source	Benchmark	Large 4 result	Comparison point	Context
Hugging Face 1	DeepSWE v1.1	61.7%	—	Reported score
Hugging Face 1	SWE-Atlas-QnA	59.4%	—	Reported score
Hugging Face 1	Terminal-Bench 4	28.3%	—	Reported score
Hugging Face 1	Combined Coding Agent Index	49.8%	Ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max	Reported ranking
ZDNet 2	Terminal-Bench 4.0	41%	Up from 28% at preview; one point above GLM-5.3	ZDNet’s comparison
ZDNet 2	Coding Agent Index	48	GLM-5.3 at 49; Codex GPT-6.1 Terra at 48	ZDNet’s comparison
Hugging Face 1	Artificial Analysis Cyber Index	82%	—	Tests reproducing and patching a real open-source vulnerability

Hugging Face reports 28.3% on Terminal-Bench 4, while ZDNet reports 41% on Terminal-Bench 4.0. Hugging Face gives a Coding Agent Index score of 49.8%; ZDNet reports 48. ¹² The Terminal-Bench figures differ by 12.7 percentage points. Neither source explains whether differences in checkpoints, prompts, harnesses, reasoning modes, sampling settings, or test dates account for the results. ¹² Until the discrepancy is clarified, treat these as source-specific reports rather than a settled result.

Hugging Face says Large 4 ranked fifth among open-weight models on the Artificial Analysis Intelligence index as of October 29, 2026. ¹ ZDNet reports Mistral’s claim that Large 4 used 47% fewer tokens per task on average than GLM-5.3 on the Coding Agent Index. ² Token use alone does not establish total serving cost or performance on a buyer’s own tasks.

What the Mistral Large 4 benchmark results skip over

A benchmark result is evidence about a test, not a production guarantee. Hugging Face says its Cyber Index test asks models to reproduce and patch a real vulnerability in open-source software. ¹ ZDNet describes CyberGym as testing whether models can find and reproduce memory-safety bugs in C/C++ projects, then patch them without breaking tests; it reports Large 4 scoring just above DeepSeek V4.1 Flash. ² Those results are relevant to security work, but do not establish how the model will perform on a particular codebase or incident workflow.

Hugging Face also cites Mistral’s internal testing on malware analysis, vulnerability prioritization, and detection-rule writing. ¹ The excerpt provides limited detail about that testing, so it is not enough to treat it as independent evidence of enterprise security performance. ZDNet cautions that benchmark scores deserve scrutiny because models can game them. ²

The sources do not report reliability across repeated runs, tool-use failure rates, latency, throughput, cost per completed task, or performance on private code and internal workflows. ¹² Analysis: a strong aggregate score can justify a pilot, but it does not replace testing the full workflow, including failures and retries. For a related look at how a headline coding-agent score can leave harder tasks unsolved, see InfraDev Tests Coding Agents on Multi-GPU Infrastructure. The Best Score Is 38.9%.

Hardware and cost: the deployment math is still missing

The reported one-trillion total and 49-billion active parameter counts describe model scale, not the memory, hardware, batching, or serving configuration needed to meet a particular latency or throughput target. ¹ The sources do not state those requirements for the FP8, NVFP4, or community GGUF versions, or quantify quality, speed, and resource trade-offs between them. ¹

ZDNet reports Mistral’s statement that it trained the model from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months and deployed it in its own data centers in Europe. ² That provides training and deployment context for Mistral, not an inference recipe for another organization’s infrastructure.

Analysis: test candidate variants under the conditions you expect to serve: representative prompts, context lengths, concurrency, tool calls, and image inputs. Track task success and failure alongside latency, resource use, and cost. Compare self-hosting with hosted alternatives on total operating cost, including hardware, serving engineering, monitoring, storage, upgrades, support, and the cost of failed or retried tasks. For the hosted side, a pricing page that shows final prices makes the per-task arithmetic easier to check. The sources do not provide a complete cost model for either path. ¹²

Open-weight license and service model need separate review

Hugging Face identifies the license as the Mistral Ecosystem Open-Weight License, an MIT derivative that establishes limitations for model-as-a-service offerings. ¹ That description does not settle whether a specific commercial architecture is permitted. Legal teams should review the license text against the planned deployment.

Weights and managed inference are separate decisions. Hugging Face lists inference endpoints alongside local tooling, but the excerpt does not provide endpoint access terms, availability, support arrangements, uptime figures, or service-level commitments. ¹ Treat licensing, security review, data handling, and operational support as separate approval tracks; benchmark scores do not answer those questions. Teams that put an enterprise AI gateway in front of their models can add routing constraints such as region, data policy, and allowed providers as a further control, but that does not replace the license review.

What to run before adopting Mistral Large 4

Build a workload matrix around the reason for evaluating the model: coding agents, vulnerability analysis, multimodal input, long-context retrieval, or ordinary business prompts. Mistral positions Large 4 for coding, cybersecurity, and multimodality, but that positioning does not establish fit for a specific buyer’s work. ¹ Use representative private data only under your organization’s approved testing controls.

Run identical prompts and tool harnesses against Large 4, relevant hosted alternatives, and at least one other weight-accessible model. A multi-provider model catalog can help line up the hosted candidates for that comparison. Track task success, patch validity, regression failures, latency, token use, retries, and cost per successful task; one aggregate score will not capture all of these dimensions.

Before using either Terminal-Bench result in a procurement decision, identify the model version, reasoning mode, prompt set, harness, sampling settings, and hardware behind it. The sources do not resolve the gap. ¹² Then test the chosen variant at the context length and concurrency you need, and complete licensing and operational review before deployment.

Large 4 merits a shortlist slot when weight access is a requirement. The reported rankings are a reason to test it, not proof that it will win for every workload, deployment model, or budget. ¹² An independent rerun of Terminal-Bench with a published harness would help clarify the reported results.