Mistral Large 4 Open-Sources 1 Trillion Parameter Weights, Leading Open-Source Models on Coding Benchmarks

Mistral officially opened the weights of the Large 4 model in October 2026. The model has a total parameter scale of 1.05 trillion, adopts a mixture-of-experts design, activates approximately 52 billion parameters during inference, and is equipped with a 160 million parameter vision encoder to process image inputs.

According to the official blog, three weeks before this weight release, the model was optimized through reinforcement learning and achieved a score of 19.7% on the Cognition FrontierCode 2.0 benchmark, higher than Beam 1.0's 2.9% and GLM-5.3 Flash's 14.2%.

Practical deployment differences brought by training on European infrastructure

The model was trained from scratch in European data centers, and its parameter lineage is clearly traceable, allowing enterprises to deploy it locally without relying on external API call chains.

Data shows that the model supports a 1 million token context window and has recorded 19.3K downloads, proving that developers have begun to actually download and run it.

Models trained in Europe provide a controllable foundation for sovereign AI scenarios, and institutions do not need to send code and security data to overseas servers.

At the same time, the model performs prominently on defensive cybersecurity tasks, while some Chinese open-source models still maintain a local lead on terminal command processing benchmarks.

The direct impact of improved inference efficiency on cost structure

The reinforcement learning stage significantly improved inference token efficiency, and the number of tokens consumed for the same coding task is lower than that of dense models of comparable scale.

This efficiency directly translates into lower hardware costs in self-hosted scenarios, and developers can choose between performance and cost across three inference modes.

Recorded API pricing shows $0.70 per million input tokens and $2.10 per million output tokens; after the weights are opened, local deployment can completely avoid these fees.

The execution path from preview to weight release

The model was first made available in preview through the Mistral Studio API, and the weight files were delivered later in the same month, with the actual delivery time consistent with the original commitment.

Reports noted that the preview stage already supported multimodal and agentic coding workflows, and after the weights were opened, users could directly run the same capabilities locally.

This pace shows that Mistral remained consistent between feature promises and actual delivery, avoiding the situation of releasing only a demo while delaying the weights.

Changes in parameter utilization under the open-source competitive landscape

Of the 1.05 trillion total parameters, only about 5% are activated for inference. This sparse design reduces the computation required for a single forward pass while maintaining frontier coding performance.

Feedback from the developer community shows that the model has entered a practically deployable stage in agentic coding and cybersecurity scenarios, rather than remaining only a benchmark demonstration.

Multilingual capability is stable across major languages, but gaps remain in low-resource languages. This limitation can be gradually alleviated through community fine-tuning after the weights are opened.

Independent media have reported the above benchmark results, and the factual basis is reliable.

Independent judgment

The core value of open-sourcing the Mistral Large 4 weights lies in providing coding capabilities at the 1.05 trillion parameter scale in a locally runnable form to developers in Europe and around the world, which will drive faster iteration of self-hosted agent tools in the short term.

In the long term, the combination of the European training path and open weights may form another branch of technological evolution running in parallel with the closed-source route in North America.