Mistral AI has released Mistral Large 4 in public preview. According to the company's newly published documentation, the update is an open-weight, multimodal model built on a granular Mixture-of-Experts architecture. It features 1.05 trillion total parameters—with 49 billion active during inference—a 1.6-billion parameter vision encoder, and a massive 1-million token context window.
While comprehensive release notes and official benchmark charts are limited so far, the published technical specifications outline a major frontier model update.
Core Specs and Pricing
Mistral has positioned Large 4 with aggressive pricing for its capabilities. Input tokens cost $0.68 per million, which drops to $0.07 per million for cached inputs, while output tokens are priced at $2.09 per million.
Crucially for automation builders, the documentation confirms the model supports a suite of essential operational features. This includes structured outputs, function calling, document Q&A, prefix support, batching, and built-in tools for agents and conversations.
What This Means for SMB Automations
For small and mid-sized businesses automating their administrative and operational workflows, Mistral Large 4 is a compelling addition to the toolkit.
First, the combination of structured outputs and function calling means this model is built to act as a reliable routing agent within business workflows. It can parse unstructured data and reliably trigger external software actions. When paired with the 1-million token context window, businesses can feed entire employee handbooks, months of customer service logs, or dense financial documents into a single prompt for analysis without losing the thread.
Second, as a high-tier open-weight model from a major European lab, it offers a vital alternative to the dominant American and Chinese models. For businesses operating in regulated industries or dealing with sensitive client data, having a highly capable open-weight model that can be deployed in tightly controlled environments provides a critical safeguard for data privacy.
Even for companies relying entirely on API access, the aggressive input caching costs ($0.07 per million tokens) make it economically viable to run repetitive, high-context administrative tasks. As the market digests the model's real-world performance, Mistral Large 4 ensures SMBs have more leverage to balance cost, privacy, and capability in their automation stacks.