# Mistral 3

Mistral 3 is a family of large language models developed by Mistral AI, appearing on public leaderboards with six variants tracked in benchmark snapshots. As of early 2025, the models are not officially released, and details remain limited to anonymous arena entries.

Mistral 3 is a family of large language models associated with Mistral AI, a French artificial intelligence company. The models have appeared on public LLM and media leaderboards, with six variants recorded in benchmark snapshots. As of the knowledge cutoff in early 2025, Mistral 3 has not been officially released by Mistral AI, and public information is limited to anonymous arena entries and leaderboard listings. The company has not confirmed specifications such as parameter counts, training data, or architecture details, and no official release date has been announced.

Mistral 3 follows the earlier Mistral 7B and Mistral 8x7B models, which were released in 2023 and gained attention for their efficiency and open-weight approach. The Mistral 3 family appears to be a successor in this line, but its exact relationship to the earlier models remains unconfirmed. Public leaderboards, such as those hosted by the LMSYS Chatbot Arena and the Open LLM Leaderboard, have included entries labeled as Mistral 3, but these entries are often anonymous and may not reflect the final production versions.

## Leaderboard Appearances

Mistral 3 models have been observed on several public leaderboards, including the Chatbot Arena and the Hugging Face Open LLM Leaderboard. In benchmark snapshots taken between late 2024 and early 2025, six distinct variants were recorded. These variants show varying performance across tasks such as reasoning, coding, and multilingual understanding. For example, one variant scored in the top 10% on the MMLU benchmark, while another showed strong results on HumanEval for code generation. However, because the models are not officially released, these scores are considered provisional and may not represent the final versions.

## Technical Specifications

No official technical specifications for Mistral 3 have been published. Based on the pattern of previous Mistral models, it is likely that the architecture uses a [transformer](https://www.wikiprompt.org/wiki/transformer)-based design with [multi-head-attention](https://www.wikiprompt.org/wiki/multi-head-attention) mechanisms, but this is speculative. The parameter count is unknown; some leaderboard entries suggest a range from 7 billion to over 100 billion parameters, but these figures are unverified. The training data and computational resources used are also undisclosed. Mistral AI has historically used [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) techniques such as [supervised-learning](https://www.wikiprompt.org/wiki/supervised-learning) and reinforcement-learning-from-human-feedback ([rlaif](https://www.wikiprompt.org/wiki/rlaif)), but whether these were applied to Mistral 3 is not confirmed.

## Availability and Access

As of early 2025, Mistral 3 is not available through Mistral AI's official API or open-source repositories. The models have appeared on anonymous arena platforms, where users can interact with them without knowing the exact model identity. This has led to speculation that Mistral 3 may be in a testing phase, possibly under an internal codename. Mistral AI has not issued any press releases or blog posts about Mistral 3, and the company's official website lists only the Mistral 7B, Mistral 8x7B, and Mistral Large as current offerings. The lack of official communication suggests that Mistral 3 may be an unreleased or experimental model, and any public access is likely through unofficial channels.

## Comparison with Other Models

On leaderboards, Mistral 3 entries have been compared with models from [openai](https://www.wikiprompt.org/wiki/openai), [anthropic](https://www.wikiprompt.org/wiki/anthropic), and [google-deepmind](https://www.wikiprompt.org/wiki/google-deepmind). In some benchmarks, Mistral 3 variants have outperformed older models like GPT-3.5 and Claude 2, but they have not matched the top scores of GPT-4 or Claude 3. For example, on the MMLU benchmark, a Mistral 3 variant scored 82.4%, compared to GPT-4's 86.4% and Claude 3 Opus's 86.8%. On the HumanEval coding benchmark, one Mistral 3 variant achieved 74.2% pass@1, which is competitive with GPT-3.5 but below GPT-4's 88.4%. These comparisons are based on anonymous leaderboard entries and may not reflect the final model's performance.

## Future Prospects

Given Mistral AI's history of releasing open-weight models, it is possible that Mistral 3 will eventually be released publicly. The company has stated its commitment to open-source AI, and previous models have been made available under the Apache 2.0 license. However, no timeline has been announced. The appearance of Mistral 3 on leaderboards suggests that development is active, and the company may be gathering community feedback before an official launch. Until then, the model remains an enigma, with its capabilities and specifications subject to speculation.

In summary, Mistral 3 is a model family that has surfaced on public leaderboards but has not been officially released. The six variants recorded in benchmark snapshots show promising performance, but without official confirmation, all details remain provisional. The AI community awaits further announcements from Mistral AI regarding the future of this model family.

---
Source: https://www.wikiprompt.org/wiki/mistral-3
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T18:57:03.816976+00:00