# Mistral Large 4 puts a trillion-parameter coding model in your hands

Mistral released downloadable Large 4 weights with a 1M-token context window, three reasoning modes and a custom license. VentureBeat reports stronger efficiency, while Mistral and Hugging Face document the downloadable checkpoints and custom license.[1][2][4]

Mistral Large 4 is now available as FP8 and NVFP4 weights, through Mistral's API and on partner platforms.[1][2] The sparse multimodal model has about one trillion total parameters and 49 billion active parameters per token.[2] Mistral positions it for coding and defensive security, while independent evaluation places it second among current open-weight coding-agent configurations and well behind the leading closed models on broader capability.[4] Teams should test the full serving and agent stack on their own work before moving sensitive code or security workflows.

## What changed on October 29

[Mistral released the improved Large 4 weights [1]](https://mistral.ai/news/mistral-large-4-oss/) after an API preview earlier in October.[1] The release adds low, medium and high reasoning modes, downloadable FP8 and NVFP4 checkpoints, and deployment recipes for vLLM and SGLang.

[Hugging Face's launch guide [2]](https://huggingface.co/blog/mistral-large4) describes a natively multimodal mixture-of-experts model with about one trillion total parameters, 49 billion active parameters per token and a one-million-token context window.

[Reuters confirmed the public release and its cyber focus [3]](https://www.reuters.com/legal/litigation/mistral-ai-publicly-releases-large-4-targets-cybersecurity-users-2026-10-29). Anyone who accepts the license can download the model.[1][5] Organizations gain control over deployment, and they also assume responsibility for access controls after download.

## How to read the benchmark results

Mistral reports 48 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than GLM-5.3.[1] The company calls the results effectively tied and also reports 41% on Terminal-Bench 4.0 after three more weeks of reinforcement learning.[1]

[An independent benchmark review [4]](https://venturebeat.com/technology/mistral-large-4-le-chonk-arrives-as-the-top-open-model-outside-china-with-improved-coding-performance-and-63-lower-cost-per-task) places Large 4 second on the coding index, one point behind GLM-5.3. It also reports a broader Intelligence Index score of 26, six points behind the leading open-weight models and far below the leading closed systems.

The reported coding scores cover complete agent configurations, including the harness and tools, so each score measures more than the model. Mistral's Terminal-Bench result and the evaluator's separate result also use different configurations, so neither number should be treated as a universal completion rate.[4]

## A practical evaluation plan

- Choose the exact checkpoint, quantization, serving engine, reasoning mode and agent harness. Record them with every result.
- Build a set of representative repository tasks with known acceptance criteria, regression tests and human-reviewed outcomes.
- For security work, separate vulnerability discovery, reproduction and patching. Run exploit validation only against owned or explicitly authorized isolated systems.
- Compare Large 4 with your current model using accepted changes, valid findings, false positives, regressions, wall-clock time, total tokens and reviewer hours.
- Test long-context retrieval with realistic repositories. A one-million-token limit does not prove that the model will find or use the right evidence.
- Seed instructions that should not be followed in issues, docs and dependencies, then verify that the agent keeps its scope and protects credentials.
- Set limits for network access, secrets, tool permissions, spend and run time before testing high reasoning mode.
- Repeat the test across several runs. Agent results vary, and one successful demonstration does not establish a reliable workflow.

## License, price and deployment choices

[The Mistral Ecosystem Open-Weight License [5]](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE) permits use, modification, fine-tuning and commercialization, but model-as-a-service businesses need a separate commercial license.[1][5] Legal review should confirm that your distribution and hosted-service plans fit the terms.

Mistral lists API rates of $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.[1][6] Self-hosting replaces that bill with accelerator, storage, networking, operations and utilization costs.

The official release points to Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted options in addition to downloadable weights.[1] Compare data handling, model version control, observability, exit costs and total operating cost across the available deployment choices.

## Set tighter boundaries for cyber work

Mistral reports 51% on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on its malicious cyber request set.[1] It also says live traffic showed possible uplift for real attacks and that it is modifying guardrails.[1]

Downloaded weights let operators alter or remove hosted controls. Security teams gain deployment control, but they also become responsible for identity, network limits, audit logs, incident response and abuse monitoring.

Teams should not test a live third-party system on the strength of a defensive benchmark. Keep targets explicit, require human approval before exploitation and preserve evidence for every command and patch.

## What remains uncertain

- The strongest launch scores come from Mistral or from configurations that combine a model with a particular harness.[1][4]
- Mistral reports 97.1% resistance on the B3 agent-security benchmark, but that result does not cover every repository, tool or prompt-injection path.[1]
- Teams still need to measure hardware cost and latency for their chosen checkpoint, quantization and workload.
- VentureBeat reports a smaller broad capability gain than Mistral’s launch language suggests and a larger improvement in efficiency.[4]
- The custom license permits many uses, differs from the standard MIT license and includes a model-as-a-service condition.[5]
- Mistral says it is changing cyber guardrails after observing possible attack uplift, so hosted behavior may change after publication.[1]

## Frequently asked questions

- Are the weights available now? Yes. Mistral links FP8 and NVFP4 checkpoints on Hugging Face.[1][2]
- Is Mistral Large 4 open source? Mistral calls it open weight. The weights use a modified MIT license with a separate condition for model-as-a-service businesses.[1][5]
- Is it the best open coding model? Mistral presents it as the leader, while independent results place its tested agent configuration one point behind GLM-5.3.[1][4]
- Can it run on one consumer GPU? The official full checkpoints target multi-accelerator serving. Test a supported quantization against your hardware and quality requirements.[2]
- Should a team replace its current coding model? Run a matched evaluation first. Compare accepted work, cost, latency, review time and security failures.
- Can it test production systems for vulnerabilities? Only with explicit authorization and strict containment. Start with isolated repositories and challenge targets.

## Sources

- [[1] Mistral: Large 4 open-weight release](https://mistral.ai/news/mistral-large-4-oss/)
- [[2] Hugging Face: Mistral Large 4 launch guide](https://huggingface.co/blog/mistral-large4)
- [[3] Reuters: public release and cybersecurity focus](https://www.reuters.com/legal/litigation/mistral-ai-publicly-releases-large-4-targets-cybersecurity-users-2026-10-29)
- [[4] VentureBeat: independent benchmark analysis](https://venturebeat.com/technology/mistral-large-4-le-chonk-arrives-as-the-top-open-model-outside-china-with-improved-coding-performance-and-63-lower-cost-per-task)
- [[5] Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE)
- [[6] Mistral Large 4 model documentation](https://docs.mistral.ai/models/mistral-large-4-0)