Mistral reportedly open-sources improved Large 4 weights

AT A GLANCE

According to the report, Mistral open-sourced Large 4 weights; programming performance improved, but third-party retesting showed discrepancies.

Article

Mistral open-sources trillion-parameter Large 4 weights: programming self-test rises by 13 percentage points, per-task cost falls by 63%

Beating AI Flash: France's AI company Mistral officially open-sources the weights of its flagship model Large 4(Le Chonk). The model has total parameters of 1050000000000, with approximately 49000000000。October 6 activated per inference.API Following the open 3 preview, Mistral conducted another weeks of reinforcement learning training; this release is an improved version.

Mistral's announced Terminal-Bench 4.0 programming score rose from 28%to 41%；FrontierCode 2.0 with a score of 19.7%, ranking first among open-source models. In Artificial Analysis's coding Agent index, Large 4 scored 48%, close to GLM-5.3's 49%, while consuming on average 47%less Token per task. However, under its own test configuration, Artificial Analysis measured Terminal-Bench 4.0 at only 25%, failing to reproduce the official 41%。

In other tests, the financial spreadsheet task Finch rose from 67.4%to 73.3%, and the agent's success rate in resisting prompt-injection attacks improved from 93.3%to 97.1%. Its cybersecurity index score was 51%, still below DeepSeek V4.1 Flash's 57%。

The improvement in overall capability was less substantial. Artificial Analysis has just upgraded its intelligence index to v5; on a comparable basis, Large 4 rose from 25 points in the preview version to 26 points, below GLM-5.3 and DeepSeek V4.1 Flash's 32 points.

However, inference efficiency improved significantly: output per task Token decreased by 39%, and the hallucination rate fell from 41%to 28%。API. Pricing was also nearly halved. Based on October 29's evaluation pricing and usage estimates, per-task cost fell from 5.54 USD to 2.06 USD, a reduction of 63%。

The model weights have been uploaded to Hugging Face, with FP8 and NVFP4 versions available, supporting self-deployment, fine-tuning, and general commercial use. Operating a commercial model-hosting service still requires separate authorization from Mistral.

Original link https://m.theblockbeats.info/flash/374669

Key points

Mistral released the model weights for Large 4(Le Chonk), with total model parameters of 1050000000000, and approximately 49000000000 parameters activated per inference.
The report states that, after Mistral's October 6 open API preview, it conducted another 3 weeks of reinforcement learning training; this release is an improved version.
Mistral's announced Terminal-Bench 4.0 programming score rose from 28%to 41%；FrontierCode 2.0 with a score of 19.7%, ranking first among open-source models.
Artificial Analysis measured Large 4 at 48%in its coding Agent index, consuming on average 47%less Token per task; however, under its test configuration, Terminal-Bench 4.0 scored 25%, failing to reproduce the official 41%result.
The financial spreadsheet task Finch rose from 67.4%to 73.3%, and the success rate in resisting prompt-injection attacks rose from 93.3%to 97.1%。
Large 4. The cybersecurity index score was 51%, below DeepSeek V4.1 Flash's 57%。
Artificial Analysis's intelligence index v5 On a comparable basis, its scoring shows that Large 4 rose from 25 points in the preview version to 26 points, below GLM-5.3 and DeepSeek V4.1 Flash's 32 points.
The report states that output per task Token decreased by 39%, and the hallucination rate fell from 41%to 28%。
Based on October 29's evaluation pricing and usage estimates, per-task cost fell from 5.54 USD to 2.06 USD, a reduction of 63%。
The model weights have been uploaded to Hugging Face, providing FP8 and NVFP4 versions that support self-deployment, fine-tuning, and general commercial use; commercial model-hosting services still require separate authorization from Mistral.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model with a very large parameter scale. Reports say it has released downloadable weights, allowing users to deploy or fine-tune it themselves. Some coding and efficiency metrics have improved, but results from different testing sources are inconsistent, and the overall intelligence score has improved only marginally.

Why it matters to readers

If the weights and licensing information are accurate, developers may reduce their reliance on API and attempt self-deployment, fine-tuning, or lower inference costs.

You can think of it as a model version that can be downloaded and run;AI however, the licensing terms for downloading, deployment, and commercial hosting are not the same, so one cannot assume that all uses are free based solely on the term “open source.”

Risks and unknowns

The evidence comes from media reports and does not provide an official Mistral announcement as primary proof.
Artificial Analysis's Terminal-Bench 4.0 result is 25%, which is inconsistent with the 41% announced by Mistral.
The availability of the weights, model version, and specific licensing terms still need to be confirmed by consulting Mistral and Hugging Face pages.
The cost per question is based on October 29 estimated evaluation pricing and usage; actual costs may differ.

Mistral reportedly open-sources improved Large 4 weights

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center