Mistral AI opened public access to Mistral Large 4 on Tuesday, putting a one-trillion-parameter model behind its API now and promising downloadable weights on October 27. The Paris company is pitching the model as a European answer to the open-weight systems that Chinese labs have pushed to the front of the market.
The public preview arrives with a one-million-token context window, text and image input, and support for function calling, structured output, document Q&A, batching, agents and built-in tools. Mistral’s model documentation lists 1.05 trillion total parameters, 49 billion active parameters and a 1.6-billion-parameter vision encoder.
Those specifications make Large 4 important, but not yet easy to judge. Mistral is still training the release checkpoint, its headline benchmark results have not all been reproduced independently, and the license covering the weights has not been published. For developers and enterprise buyers, the next three weeks matter as much as launch day.
Mistral Large 4 is huge, but it does not run all trillion parameters at once
Large 4 uses a mixture-of-experts architecture. A routing system selects a subset of specialized parameter groups for each token instead of activating the entire network for every request. That is why the model can contain 1.05 trillion parameters while using 49 billion at inference time.
The distinction affects both capability and deployment planning. Total parameters describe the model’s capacity, but active parameters are a better clue to the computation required for each generated token. Neither number, on its own, tells a customer how much GPU memory, interconnect bandwidth or serving infrastructure a self-hosted installation will need. The complete weights still have to be stored, and large sparse models can be demanding to distribute efficiently across accelerators.
Mistral trained the model from scratch in about two months on roughly 4,000 Nvidia Grace Blackwell GPUs in its European data centers, according to interviews given to Axios and VentureBeat. The company says its training corpus covers more than 160 languages, including every official language of the European Union.
The API produces text, even when a prompt includes images. Mistral is highlighting software engineering, cyber defense, financial analysis, technical drawings and satellite imagery as target workloads. Large 4 is also expected to become the default model in Mistral’s Vibe coding assistant.
The API price is clear; the self-hosting economics are not
Mistral’s documentation currently displays public-preview pricing of $0.68 per million input tokens, $0.07 per million cached input tokens and $2.09 per million output tokens. That is inexpensive for a flagship-scale API, although token prices alone do not capture latency, tool-call reliability, long-context accuracy or the amount of reasoning a model needs to finish a task.
The one-million-token context window is useful for large repositories, document collections and lengthy agent histories, but it also changes the cost calculation. A request containing 500,000 uncached input tokens costs about 34 cents before output. Reusing the same 500,000 tokens from cache would cost roughly 3.5 cents. Teams evaluating Large 4 should therefore measure cache-hit rates and end-to-end task cost, rather than comparing only the advertised price for a million tokens.
Self-hosting is a separate question. Mistral has not yet released memory requirements, recommended quantizations, throughput measurements or a reference deployment topology for Large 4. A 49-billion-active-parameter model may be computationally sparse, but a one-trillion-parameter checkpoint will still sit well outside the hardware budget of most individual developers and smaller companies. The October 27 release should show whether Mistral supplies smaller precision formats and practical serving recipes.
Mistral’s benchmarks look competitive, not conclusive
Mistral reports a score of roughly 62% to 63% on DeepSWE 1.1, a benchmark for long-horizon software-engineering work. That would put Large 4 close to strong Chinese open-weight systems under the comparison configuration used in its launch material.
The ranking changes when different agent harnesses and published runs are compared. VentureBeat noted that the live DeepSWE leaderboard shows leading model-and-agent combinations near 69% for GLM-5.3 and Kimi K3, while top closed systems reach about 74%. Large 4 had not appeared in the public leaderboard or in Artificial Analysis testing at publication time.
Other launch results are promising but carry similar limits. Mistral reports a 15% task-pass rate on the Harvey Legal Agent Benchmark and 67% on Finch, which tests finance and accounting workflows built around documents, spreadsheets, search and reporting. It also claims strong visual grounding on Dense200 and DIOR-RSVG. Some competitor figures in the supplied comparisons can be traced to public leaderboards; others have not been published under matching configurations.
That does not make the results meaningless. They suggest that Large 4 deserves independent testing across coding, finance and visual tasks. They do not yet prove that it is the best open-weight model overall, or that a benchmark advantage will survive different prompts, tools and inference settings.
“Open weight” will depend on the October 27 license
Mistral is making the model available first through a moderated API. A version with broader cybersecurity capabilities is being shared with selected partners while the company completes additional reinforcement learning and safety testing. The planned weight release follows on October 27.
The company calls Large 4 open weight, but that term describes access to the model files, not the legal permission to use them. VentureBeat reported that the weights are expected under a custom Mistral license. Mistral’s existing catalog spans Apache 2.0 models and models governed by a modified MIT license that requires companies above a revenue threshold to obtain a commercial license or use Mistral’s hosted service. Its license guidance tells users to consult each model card for the applicable terms.
Until Large 4’s model card and license are public, organizations should not assume unrestricted commercial deployment, redistribution or derivative-model rights. “Downloadable” and “open source” are not interchangeable.
What developers should test now
Teams can use the preview to establish a baseline before committing to a migration. A useful evaluation should include the actual tools, prompts, repositories and document sets used in production, with particular attention to five areas:
- Long-context retrieval: Test whether the model finds relevant information throughout a large prompt, rather than only accepting a million tokens.
- Agent reliability: Measure tool-selection errors, malformed arguments, recovery after failures and the number of steps needed to finish a task.
- Multilingual quality: Evaluate the languages and regional terminology that matter to the deployment instead of treating the 160-language claim as uniform performance.
- Cost and latency: Record uncached and cached input, output length, time to first token and total task completion time.
- Safety and control: Check how the moderated API handles legitimate security work, sensitive documents and requests that sit near policy boundaries.
The weight release should trigger a second round of testing focused on license terms, quantization quality, hardware fit, throughput, isolation and security controls. Mistral Large 4 is a credible new option today, but the evidence needed for a production self-hosting decision will arrive later than the launch headline.
Sources: Mistral model documentation, Axios, VentureBeat, Le Monde, and Mistral licensing guidance.
