Mozilla AI Adds Mistral Large 4 to Otari With Existing Controls Intact

Mozilla AI is promising a simpler model switch, not a proven performance upgrade. On October 8, it added Mistral Large 4 to Otari, its open-source AI gateway. Existing users can switch with one model-string change while retaining their policies, request records and cost tracking—but Mozilla AI recommends testing performance on their own workloads.

Otari sits between applications or AI agents and the models and tools they call. Mozilla AI describes it as provider-agnostic: routing, policy enforcement and activity tracking live in that shared layer rather than a particular model provider’s console. Large 4 joins the other Mistral models already available through the gateway.

A model selector, not a new integration

Models use a provider-and-model selector, so an existing Otari integration can point at Large 4 by changing that selection. Provider credentials are managed centrally. The gateway also handles routing and provider failover, letting teams run Large 4 alongside models from other providers without distributing a new set of keys across their services.

New users can create an API token on the hosted Otari platform and install its Python software kit. Applications already using the OpenAI software kit can keep that code and change the base URL to Otari’s endpoint. Mozilla AI also offers an open-source repository for teams that want to self-host the gateway.

The rules stay at the gateway

Mozilla AI warns that swapping an agent’s model can change which tools it chooses and how it behaves. Guardrails tied to one provider’s platform do not carry over to another. Otari’s approach is to keep those policies outside the model provider, applying them when traffic reaches Large 4.

Otari connects to agent hooks—the points before an agent processes a prompt or runs a tool. It evaluates two kinds of checks at those points, allowing agents to act within boundaries set by the team without requiring a person to approve every step.

• Rule-based checks inspect recorded evidence: whether a protected file was edited, a forbidden command was attempted or required tests passed.
• AI judgment checks assess less clear-cut requirements, such as whether a change follows the team’s architecture or meets its acceptance criteria.

Model calls, agent actions and policy decisions enter one history with timestamps and per-request costs. That creates a shared record for comparing models, rather than leaving teams to piece together activity from separate provider systems.

Gateway access is not a weights release

Mistral’s model documentation lists 1.05 trillion total parameters and a 1.6-billion-parameter vision encoder. The company pitches Large 4 for tasks combining visual inspection and tool use, including analyzing engineering drawings, complex documents and satellite imagery. Those are Mistral’s capability claims, not results demonstrated by the Otari integration.

The companies give different figures for the active portion of its mixture-of-experts design, which uses only part of the model for each token. Mozilla AI lists 49 billion active parameters; Mistral’s documentation and launch announcement list 52 billion. Neither explains the difference. The total parameter count should not be confused with the amount activated during a response.

Mistral says it trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers, which also serve the public preview. Ahead of the weights release, it says cybersecurity leaders, vetted partners and state authorities are red-teaming the model with reduced moderation and expanded cyber capabilities.

Mistral opened the public preview on October 6, 2026, and said downloadable weights would follow by the end of October. Self-hosting Otari therefore should not be confused with self-hosting Large 4: the gateway’s deployment options and the model’s planned weights release are separate.