Mistral Large 4 Preview scores 38 on intelligence index

Artificial Analysis reported on October 19 that Mistral Large 4 Preview scored 38 on its intelligence index (version 4.3.2), while Meta's Muse Spark Open scored 39. The results come from a comparison of several large language models, including DeepSeek V4.1 Flash, which also scored 39.

Mistral Large 4 is a preview version

Mistral Large 4 is a preview version, and its reinforcement learning is still in development, according to Mistral's announcement. The model has not yet been fully released, and its weights are not publicly available. This means it is still classified as a proprietary model by Artificial Analysis, even though it is being evaluated for open model comparisons.

Mistral Large 4 outperforms Muse Spark Open in agent tests

In agent-based tests such as Terminal-Bench and AutomationBench, Mistral Large 4 outperformed Muse Spark Open. On Terminal-Bench 4.0, Mistral Large 4 achieved a 27% success rate compared to Muse Spark Open's 8%. On AutomationBench-AA, Mistral Large 4 achieved 60% success compared to 50% for Muse Spark Open.

Muse Spark Open leads in knowledge and science tests

Muse Spark Open outperformed Mistral Large 4 in knowledge and science tests. It scored higher in Humanity's Last Exam (44% vs. 35%), SciCode (58% vs. 54%), and AA-Omniscience (25 vs. −5). The AA-Omniscience test penalizes incorrect answers, and a negative score indicates more incorrect than correct answers.

Mistral Large 4 is more verbose and costly

Mistral Large 4 produces significantly more output tokens than Muse Spark Open, with 200 million tokens compared to 110 million. This higher verbosity increases the cost of running the intelligence index by 36%, from $1,183 for Muse to $1,607 for Mistral Large 4. The execution time for Mistral Large 4 is also longer, taking 734 seconds compared to 288 seconds for Muse.

Mistral Large 4 may soon be open-sourced

Mistral AI plans to release the weights for Mistral Large 4 by the end of October. The company has indicated that the weights will be made available on Hugging Face, with an estimated release date of October 31, 2026. However, the license for the model has not yet been published, and details about usage restrictions remain unclear.

Muse Spark Open has a commercial use license

Meta has published the license for Muse Spark Open, which allows commercial use with certain restrictions. The license requires users to notify Meta and sign a separate agreement if they plan to offer the model as a service beyond $20 million in revenue. It also mandates that any product or service using Muse Spark Open must visibly display the phrase "Built with Muse Spark Open."

DeepSeek V4.1 Flash is a strong competitor

DeepSeek V4.1 Flash, which scored 39 on the intelligence index, outperformed Mistral Large 4 on AutomationBench (69% vs. 60%) and tied with it on Terminal-Bench (27%). However, it produced 253 million tokens in the index, making it more expensive than Mistral Large 4 at the discounted price. The model is available under the MIT license, which allows more flexible usage compared to the licenses for Mistral and Muse.