Mistral Large 4 — Hosted 1M-Context Multimodal Model Playground & API

What is Mistral Large 4 (mistrallarge4.com)?

Mistrallarge4.com is an independently operated site that hosts a general-purpose multimodal model presented as Mistral Large 4, exposed through a web playground and an OpenAI-compatible REST API. Note up front: this is not an official Mistral AI property — the operator is unnamed and the footer simply reads "© 2026 Mistral Large 4" — so the model's provenance cannot be verified from the site alone.

The service targets developers and teams doing code review and debugging, long-document analysis, screenshot and diagram interpretation, agent workflows, structured data extraction, and writing or research. Its headline specification is a 1M-token context window, with the API configured around 524,288 tokens.

How Mistral Large 4 Works

The workflow has two entry points. In the playground, you sign in and interact without code: paste text or code, attach an image via a public URL, and adjust settings including reasoning effort (none or high — reasoning consumes output budget). The playground caps requests at 96 KiB with a maximum of 24 messages and 16,384 output tokens.

For integration, you create an account API key and call the OpenAI-compatible endpoint at https://mistrallarge4.com/api/v1 with model ID mistral-large-4-0, including streaming. Tool calling is supported with JSON-schema structured output, which suits agent pipelines and extraction jobs. Responses can be streamed, matching the request shape that existing OpenAI SDK code already sends — the migration is a base-URL and model-ID change.

Main Features

• 1M-token context — a claimed 1,000,000-token window, with the API configured around 524,288 tokens for long-document work.
• Multimodal input — text plus images, with images supplied as public URLs only.
• Adjustable reasoning effort — choose none or high; reasoning counts toward the output budget.
• Tool calling and structured output — JSON-schema-constrained function calls for agent and extraction pipelines.
• OpenAI-compatible API — a REST endpoint with streaming that existing OpenAI SDK code can target with a base-URL change.
• Free playground — sign in and test the model without paying, subject to request-size limits.

Pricing

The playground is free after sign-in. Beyond the free allowance, usage is charged in credits at the model's cost in USD multiplied by 12,000, rounded up. Credit packs never auto-renew but do expire 90 days after each grant: 80,000 credits cost $10; 1,000,000 credits cost $100 (marked popular, with a fast lane and 10 concurrent requests); 11,000,000 credits cost $1,000 (unlimited concurrency and a dedicated fast lane). The base tier allows 3 concurrent requests, and a monthly subscription option also exists. Exact per-call costs are hard to predict because of the multiplication and rounding formula.

Common Use Cases

• Long-document analysis — feeding contracts, reports, or books to a large context window.
• Code review and debugging — reasoning over large codebases or long threads.
• Screenshot and diagram interpretation — multimodal Q&A via image URLs.
• Agent workflows — tool calling with JSON-schema structured output.
• Structured extraction — pulling fields from messy text at scale.

Limitations

The playground accepts pasted text and public image URLs only — there are no PDF, audio, or video uploads. The site states the weights are not yet released, that they take about 0.5 TB even at 4-bit precision, and that running the model locally requires multi-GPU servers, so the hosted API is the only practical access. The playground is explicitly not a private deployment, so confidential documents should be reviewed against the privacy policy first. Credits expire 90 days after each grant, and concurrency is tiered. Most importantly, the operator is unnamed and the site is not affiliated with Mistral AI, so claims about the model's identity and benchmarks cannot be independently verified from the site.

Summary

Mistrallarge4.com packages a large-context multimodal model presented as Mistral Large 4 into an accessible free playground and an OpenAI-compatible API, with unusually concrete documentation of its own limits — input types, request caps, credit math, and hardware infeasibility of self-hosting. The value proposition is real for developers who want to test a 1M-context model cheaply. But the operator is anonymous, the site is not an official Mistral AI property, credits expire quarterly, and per-call costs are hard to predict — so it is best treated as an experimental playground to evaluate with non-sensitive data, not a production dependency.