Organisations can download and customise the text-and-image model for their own systems. Operating a model-as-a-service business requires a separate commercial licence from Mistral.
French superintelligence (SI) startup Mistral released the improved Large 4 model’s weights on 29 October 2026. These learned numerical settings let organisations download and run the model independently, rather than depend on a hosted service, as Reuters confirmed in a report carried by CNA.
Mistral says its Terminal-Bench coding-test score rose from 28% at preview to 41% at release, and that Large 4 uses 47% fewer tokens per task than GLM-5.3. These company-provided comparisons do not establish overall superiority: Fortune says leading proprietary models from OpenAI and Anthropic remain ahead in overall capabilities.
The release follows a 6 October public preview offered through Mistral’s hosted interface. Mistral says feedback from government and corporate testing, together with three more weeks of reinforcement learning—training that rewards successful task outcomes—helped improve coding and defensive-security performance before the weights became public.
Mistral reports gains in coding efficiency and resistance to attacks. The figures below are the company’s account of benchmark results, rather than independent verification of the released weights. Tokens are the small units into which text is divided for processing.
| Test | Result | Meaning or comparison |
|---|
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
| Terminal-Bench 4.0 | 41%, up from 28% at preview | A gain of 13 percentage points, with 30% fewer output tokens in terminal-based coding tasks. |
| Artificial Analysis Coding Agent Index | 48% | Effectively tied with Z.ai’s GLM-5.3, while using 47% fewer tokens per task on average, Mistral says. |
| Artificial Analysis Cyber Index | 51% | Tests finding, reproducing, validating and patching vulnerabilities without breaking existing functionality. |
| Malicious cybersecurity requests | 96.6% average refusal | Measured on cybersecurity prompts from JailbreakBench, StrongREJECT and AgentHarm. |
| B3 Agent Security Benchmark | 97.1% attack resistance | Up from 93.3% in the preview. |
The two token reductions use different comparisons. The 30% reduction concerns Large 4’s output on terminal-based coding tasks compared with its preview. The 47% figure compares average tokens per task with GLM-5.3, not with the preview. That token-use comparison is not a like-for-like measurement of competitors’ hardware requirements or total running costs.
The 82% vulnerability-reproduction-and-patching score in Hugging Face’s guide was already reported in Mistral’s 6 October preview announcement. It concerns one Cyber Index subtest, rather than the released model’s 51% overall index score.
In a testimonial published by Mistral, Brad Medairy, Booz Allen Hamilton’s president of National Cyber, said internal source-code security tests found that Large 4 matched other leading cybersecurity models’ vulnerability-detection recall—the share of flaws detected—with what he called a compelling cost curve. He said customisation and deployment under clients’ control let Booz Allen adapt the model to environments with strict rules on where data must remain.
Hugging Face says Large 4 takes text and images as input and generates text. Its model-level context limit—the amount it can process at once—is 1 million tokens. Mistral positions it for coding assistants and security work such as finding software flaws, reproducing them and writing patches.
Hugging Face reports that 2-bit quantised weights run on a single Apple Silicon Ultra computer with 512 GB of unified memory. Quantisation stores weights at lower numerical precision to reduce their size. Its separate server recipe uses eight NVIDIA H200 graphics processors, or GPUs, with FP8 weights distributed across them using vLLM model-serving software. These are configurations for particular weight formats and workloads, not universal minimum requirements.
The eight-H200 recipe supports up to 131,000 context tokens and 16 simultaneous requests. It also uses built-in draft-token generation to speed up output when those drafts are accepted. Hugging Face measured the following results; prefill means processing the input prompt, while generation means producing the answer.
| Hardware and weights | Test conditions | Measured result |
|---|---|---|
| One M3 Ultra; Q2_K quantised weights | Single-computer configuration | About 90 tokens/second prefill and 15 tokens/second generation. |
| Eight H200 GPUs; FP8 weights | One 1,000-token prompt; draft-token acceleration enabled | About 0.6 seconds to the first token and 65 tokens/second generation; 120–138 tokens/second on chat prompts where drafts were accepted. |
| Eight H200 GPUs; FP8 weights | Load test across 1,000 ShareGPT conversations | About 480 generated tokens/second in total, not per user. |
Those speeds describe Hugging Face’s tests, not a promised rate for every deployment. Its measurements on another multi-GPU setup also showed generation slowing as the context grew longer.
Mistral offers three reasoning settings so developers can choose how much work the model does before answering. The company recommends the following uses.
| Mode | Recommended use |
|---|---|
| Low | Simple tasks; replaces the former “None” mode. |
| Medium | Everyday tasks. |
| High — default | The hardest tasks requiring precision and thorough reasoning. |
Hugging Face says reasoning cannot be disabled: sending “none” as the reasoning setting still selects high.
Mistral’s model card says the Mistral Ecosystem Open-Weight License is a modified MIT licence that permits commercial and non-commercial use. Organisations can inspect, customise and deploy the model internally, but operating a model-as-a-service business—offering access to the model as a service—requires a separate commercial licence from Mistral. Downloadable weights therefore do not come with unrestricted open-source terms.
Pierre Stock, Mistral’s vice-president of science, said access cannot be contained once a model is released. That protects users from being cut off from their downloaded model, but also means anyone can use it.
The public download is distinct from a specialised cybersecurity variant that Mistral is providing in preview to enterprise design partners. That variant targets advanced security testing, including attempts to expose weaknesses by simulating attacks. Access requires vetting and contractual safeguards, and testing is restricted to owned or explicitly authorised systems in controlled environments.
For the public model, teams can use downloaded weights locally, deploy through Hugging Face’s Inference Endpoints, or use Mistral’s hosted service. Model-service businesses need the separate licence, while partners seeking the specialised security preview face its vetting conditions.
Mistral says preview monitoring showed refusals of attacks against live targets, but also potential for the model to help conduct real attacks. The company is modifying its safety controls and says it will continue updating them in the months ahead.
Stock also said smaller models are planned for the coming months, including one intended to run on a laptop, in a briefing attended by The New Stack.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min
