Mistral Large 4 goes live in Mozilla AI's Otari gateway on release

What Mistral Large 4 is

Mistral has released Mistral Large 4, which it calls Le Chonk. It is a 1.05-trillion-parameter, natively multimodal Mixture of Experts model and, per the Mozilla AI post, Mistral's largest model to date.

The model activates 49B of its 1.05T parameters per token. Mozilla AI describes this as the capacity of a trillion-parameter model at roughly the inference cost of a mid-size dense one.

On Mistral's published evaluations, Large 4 leads other US and European open-weight models on aggregate benchmarks. Mistral also reports state-of-the-art results among open models on enterprise workloads in cybersecurity, finance and manufacturing, along with strong coding and agentic performance.

Multimodality is the part Mozilla AI finds most interesting. Mistral points to complex documents, engineering drawings and satellite imagery, with the model able to inspect what it sees, reason about it and act on it inside an agentic workflow. Mistral is building the model and the infrastructure behind it in Europe, as part of a wider push toward open, sovereign AI infrastructure.

Mozilla AI is careful to note that these are vendor benchmarks and should be treated as a starting point. The numbers that matter, it says, are the ones measured on a team's own traffic.

Why a new model usually means more work

Mozilla AI argues that every frontier release widens options but also creates work in production. Adding a provider typically means another SDK, another set of API keys, different rate limits and response formats, and another invoice.

For agents the cost is higher. A model swap can change which tools an agent reaches for and how it behaves, and guardrails tied to one provider's platform do not carry over to the next. The company adds that the best-fitting model today probably will not be the best fit in six months, because costs move, capabilities shift and open-weight models keep closing the gap. In its words, most teams do not have a model problem but an infrastructure problem.

How Otari handles the new model

Otari sits between applications and agents on one side and the models and tools they call on the other. Because routing, policy and observability live in that single layer, Mozilla AI says a new model plugs into everything already running.

• Connect: Users can try the model on otari.ai, the hosted platform, by signing in and creating an API token, installing the Otari Python SDK with pip install otari, and setting the token as OTARI_API_KEY. Models use a provider: model selector, so switching means changing one string. Teams on the OpenAI SDK can keep their code and set the base URL to https://api.otari.ai/api/v1. Provider keys are managed centrally, and routing and provider failover are handled at the gateway.
• Govern: Policies live in Otari rather than in a provider's console, so they apply to Large 4 as soon as traffic is routed to it. Otari connects to agent hooks, the points where an agent is about to process a prompt or run a tool. It evaluates deterministic checks, such as whether a protected file was edited or required tests passed, and LLM-as-a-judge checks for work needing judgment, such as whether a change follows the architecture.
• Understand: Model calls, agent actions and policy decisions land in one history, with traces, timestamps and cost per request.

What it means for teams

The practical upshot, according to Mozilla AI, is that teams can validate Mistral's claims themselves. Large 4 can be compared against a current model on latency, spend and policy outcomes, using data the team already trusts.

Mistral Large 4 and the full Mistral model family are live in Otari now. Otari can be self-hosted from the open-source repository or used as a managed platform. Mozilla AI says Otari is open source and provider-agnostic, so adopting a new model does not commit infrastructure to a single vendor. The company states it is not telling users which model to pick, but building infrastructure that lets them choose, change and govern models on their own terms.