Mistral has released the full weights of Mistral Large 4, its trillion-parameter model, so any company or government can now download it, run it on its own servers and build on it. The French lab announced the release on 29 October and is pitching the model at two jobs: writing code as an agent and defending networks against cyberattacks, according to Mistral and TechCrunch.
READ ALSO: Meta's Muse Spark Open Runs a Frontier-Scale AI Model on Two Mac Studios
READ ALSO: OpenCode Breach Puts Developers' OpenAI, Anthropic and Google Keys at Risk
"A defense you don't control isn't defense," Cloudflare's Stephanie Cohen said in Mistral's launch post, summing up the case for open weights in security work.
What Mistral released
Mistral first showed Large 4 on 6 October and has now put the weights on Hugging Face. The model card lists about 1 trillion parameters in total, of which about 52 billion are active for each token, a mixture-of-experts design that keeps running costs far below what the headline size suggests. It reads text and images, handles up to 1 million tokens of context and covers more than 160 languages.
The licence is the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. Companies can use, fine-tune and sell products built on the model for free. The one exception: anyone selling the model itself as a hosted service needs a separate commercial licence from Mistral.
Developers who do not want to run it can use Mistral's API at $0.70 per million input tokens and $2.10 per million output tokens. It is also on OpenRouter, Microsoft Foundry and NVIDIA's hosted endpoints.
How good is it?
These are Mistral's own published scores, not independent tests:
| Benchmark | What it measures | Large 4 |
|---|---|---|
| Artificial Analysis Coding Agent Index | Coding as an autonomous agent | 48% |
| Terminal-Bench 4.0 | Completing tasks in a command line | 41% |
| Artificial Analysis Cyber Index | Finding, validating and patching vulnerabilities | 51% |
| BrowseComp | Finding hard facts by browsing the web | 84.5% |
| B3 Agent Security Benchmark | Resisting attacks on the agent itself | 97.1% |
Mistral says Large 4 is effectively tied with GLM-5.3, the leading Chinese open model, on the coding index while using 47% fewer tokens per task, and that it scores highest among Western open-weight models on Artificial Analysis's overall intelligence index. It does not claim to beat the best closed models from OpenAI, Anthropic or Google.
The cybersecurity bet
The security pitch is the most interesting part of the launch. Mistral says the model handles defensive work such as finding and patching bugs "without unnecessary refusals", while refusing 96.6% of clearly malicious cyber requests in its tests. Pierre Stock, Mistral's vice president of science, told TechCrunch that access to an open model "can't be cut off by definition", which matters to governments that do not want their defence tools to depend on a US company's terms of service.
Mistral is also candid about the risk. After watching live traffic, it said it sees "a potential for uplift to conduct real attacks" and is tightening its guardrails. A more specialised cyber version is in preview only for vetted partners such as Orange Cyberdefense, and only for testing systems they own or are authorised to test. Booz Allen Hamilton, Cisco, CrowdStrike, Fortinet and Zscaler are among the partners Mistral names.
What it costs to use from Nigeria
Two routes are realistic for a Nigerian developer or company, and they cost very different amounts.
Use the API. At Mistral's prices, a support bot that reads 1 million tokens of customer messages and writes 300,000 tokens of replies costs $0.70 plus $0.63, or $1.33. At the CBN rate of about ₦1,332 to the dollar, that is roughly ₦1,800, before your bank's foreign card charges. One million tokens is about 750,000 English words.
Run it yourself. This is what open weights make possible, but it is not cheap. The model card's examples spread the model across eight GPUs, and the compressed NVFP4 version targets NVIDIA's newest Blackwell chips. That is data centre hardware. For most Nigerian startups, self-hosting a model of this size makes sense only through a local cloud or data centre provider, or for a bank, telco or agency with strict data residency rules.
That second group is who this release is really for. Nigerian banks and telcos handle customer data under rules from the CBN and the Nigeria Data Protection Commission. A model they can run inside their own walls, and that cannot be withdrawn by a foreign provider, gives them an option that closed models do not.
What happens next
Mistral says its guardrail updates are ongoing, and the vetted cyber variant has no public release date. Independent benchmark results from outside groups will show whether Mistral's numbers hold up. For developers, the practical test starts now: the weights are public, and the community has already published compressed versions, including builds for llama.cpp.
By Mistral's own numbers, Large 4 is the strongest open model a Western lab has released. It is priced low on the API and free to run yourself. For African banks, telcos and governments that want AI on their own servers, it is now a serious option.
