Mistral Large 4 Enters Preview With 1 Trillion Parameters

Mistral has introduced Mistral Large 4, a new multimodal AI model with roughly one trillion total parameters and a public preview available now through the company's API. Announced on October 6, 2026, the model targets coding, reasoning, document analysis and AI-agent workflows. Its downloadable weights are not public yet.

For developers following the race between large open-weight models, that distinction is important. You can test Mistral Large 4 through a hosted API today, but self-hosting must wait for the promised weights release later in October.

Read Mistral's official announcement.

What Is New in Mistral Large 4?

Mistral describes Large 4 as a natively multimodal mixture-of-experts model, or MoE. Its launch announcement specifies about one trillion total parameters with 49 billion active parameters used during inference. The dedicated model documentation gives the more precise overall figure of 1.05 trillion parameters and lists a separate 1.6-billion-parameter vision encoder.

The company also lists a context window of up to one million tokens. That is a model capability, not a promise that every application or account can economically process a full million-token request. Long prompts can increase latency and costs, and actual supported request limits may depend on the API and service tier.

The key architectural idea is that an MoE model routes work through a subset of its parameters rather than using all of them for each token. This can improve computational efficiency compared with activating the entire network. It does not mean a model with 49 billion active parameters has the same storage or memory requirements as a 49-billion-parameter dense model. The total weight set remains enormous.

Check the official Mistral Large 4 model card.

Coding, Agents and Multimodal Workflows

Mistral is pitching the model as one system for instruction following, reasoning, code generation, tool use and visual understanding. The announcement highlights tasks such as technical-document interpretation, complex coding workflows, financial analysis and cybersecurity research.

Its developer API documentation lists support for structured outputs, function calling, document question answering and agents. Those features matter because production AI applications often need more than a fluent chat response. They need predictable JSON, tool calls, and ways to ground an answer in the documents a user supplies.

Mistral also publishes benchmark and human-evaluation results suggesting strong performance in several enterprise categories. These are useful indicators, but readers should not treat every vendor-reported ranking as an independent real-world comparison. Benchmark setup, prompting, tool access and task selection can affect the outcome.

How Much Does the Preview API Cost?

The published preview model card lists a limited two-week launch discount of 50% on token pricing. At the discounted rates shown in Mistral's documentation, input costs approximately ₱42.70 (US$0.68) per million tokens and output costs approximately ₱131.24 (US$2.09) per million tokens.

The regular listed rates are approximately ₱85.40 (US$1.36) per million input tokens and ₱262.48 (US$4.18) per million output tokens. These peso figures are indicative conversions using approximately ₱62.795 per US$1 on October 8, 2026. They are not Philippine-localized billing rates, and payment fees or exchange-rate movements may change the final amount.

Because the discount is time-limited, developers should confirm the live rate on the official pricing page before starting a large workload. Input and output tokens are billed separately, and a very long context can produce a significant bill even if the price per million tokens appears modest.

Review Mistral's current model pricing and limits.

Can You Run Mistral Large 4 on a Gaming PC?

Not yet as an officially downloadable model. Mistral says open weights are planned for release by the end of October 2026. Until then, there is no public weight package to install in local model runners or independently benchmark on consumer hardware.

Even when weights arrive, a model with more than one trillion total parameters will not become a small desktop model simply because 49 billion parameters are active per token. Quantization, expert offloading, distributed inference, storage bandwidth and the serving implementation will determine what hardware can realistically run it.

For owners of mainstream gaming GPUs with 8GB or 16GB of VRAM, a smaller model will generally be the more practical option. Anyone planning a local Large 4 deployment should wait for official files, licensing terms, memory guidance and verified community benchmarks rather than extrapolating from the active-parameter count alone.

Why This Release Matters

Mistral's announcement adds another contender to the increasingly competitive open-weight AI market. The public API offers a way to test the model now, while the promised weights could eventually give organizations more control over deployment, data residency and model customization.

There is also an important distinction between open weights and a fully open training process. Releasing parameters does not necessarily release training data, training code or every detail needed to reproduce the model.

For developers, the next milestones are straightforward: see whether the preview performs well on their own tasks, watch for the actual weights release, and compare independent performance and operating costs before making a deployment decision.

Read the Mistral launch post and the official API changelog.