Mistral Drops Large 4: A 1 Trillion Parameter Open-Weight Multimodal Giant
Mistral just released Large 4, an open-weight trillion-parameter MoE model with a 1M context window and aggressive pricing.
Models · Source: Hacker News
What happened
Mistral just launched Mistral Large 4 in public preview. It is a massive open-weight multimodal model built on a granular Mixture-of-Experts architecture. The model packs 1.05 trillion total parameters but only uses 49 billion active parameters during inference. It also includes a dedicated 1.6 billion parameter vision encoder.
The context window is a massive 1 million tokens. Mistral is pricing this aggressively for developers. Input tokens cost 68 cents per million, while cached inputs drop to just 7 cents. Output tokens are priced at 2 dollars and 9 cents per million.
The API comes fully loaded for production use. It supports structured outputs, function calling, and document QnA out of the box. Developers also get access to prefix caching, batching, and built-in agent conversation tools.
Key facts
- 1.05T — Total parameters
- 49B — Active parameters per token
- 1M — Token context window
- $0.07 — Cost per million cached input tokens
- 1.6B — Vision encoder parameters
Why it matters
Open-weight models are crossing the trillion-parameter threshold. By keeping active parameters low at 49 billion, Mistral makes it economically viable to run a massive model in production. The 1 million token context window combined with 7-cent cached inputs changes the math for document-heavy workflows. You can now stuff entire codebases or legal libraries into the prompt without burning cash.
The built-in toolset signals a shift in how model providers compete. Mistral is no longer just serving raw text generation. By integrating structured outputs, agents, and document QnA directly into the API, they are eating the middleware layer. Builders can skip third-party orchestration tools and build complex agentic workflows straight from the model provider.
For builders
Cheap context caching unlocks RAG alternatives
Cached input tokens cost just 7 cents per million. You can load massive documents once and query them repeatedly for pennies. This makes brute-force context stuffing a viable alternative to complex vector databases for many enterprise use cases.
Native agent tooling kills middleware
The API includes native endpoints for agents, conversations, and built-in tools. You do not need to rely on heavy frameworks to manage state or tool execution. Founders can ship agentic products faster with fewer dependencies to maintain.
Multimodal processing at scale
The dedicated 1.6 billion parameter vision encoder processes images alongside text. Startups building visual QA, automated inspection, or document parsing tools get state-of-the-art vision capabilities without managing a separate model pipeline.
My take
I love what Mistral is doing here. They are proving that open-weight models can scale to a trillion parameters without destroying unit economics. By giving founders frontier-level capabilities and native agent tools at commodity prices, they are making life very hard for middleware startups.
Original reporting: Hacker News. This is my rewrite and opinion.