Large 4
A trillion-parameter mixture-of-experts (MoE) large model released by Mistral under an open-weight license, supporting multimodality and a context window on the order of one million tokens.
Timeline (last 90 days)
The article notes that most of the above scores come from Mistral's own release post, and real-world performance requires further verification.
French AI company Mistral has opened the weights of its one-trillion-parameter Large 4 model for public download.
Large 4 was released in preview earlier this month, and after three more weeks of training, the improved version's weights were published on October 29.
Large 4 has image comprehension and ultra-long document reading capabilities.
The open-weight model is free for individuals, schools, and normal business use, but cloud platforms reselling the service need authorization from Mistral.
Large 4's context window reaches one million tokens and uses a sparse attention mechanism that attends to at most 2048 of the most relevant positions each time.
On the Artificial Analysis coding agent leaderboard, Large 4 scores 48%, essentially tied with the top open-source model GLM5.3, but uses 47% fewer tokens per task on average.
On TerminalBench 4.0, Large 4 scores 41%, a 13-percentage-point improvement over the preview version from three weeks earlier, and Mistral says it ranks first among open-source models.
Hugging Face testing shows that compressing Large 4 into a two-bit version requires an Apple Ultra computer with 512GB of unified memory to run.
Running Large 4 on a typical laptop with 16GB/32GB of memory generates at about 15 tokens per second.
4 more timeline events
All Facts (14)
The article notes that most of the above scores come from Mistral's own release post, and real-world performance requires further verification.
50%UnverifiedLarge 4's "thinking" feature cannot be turned off; the default is the highest setting, with low, medium, and high options available.
50%UnverifiedIn cybersecurity vulnerability reproduction and remediation tests, Large 4 scores 82%, while several open-source large models score near zero (due to outright refusal).
50%UnverifiedLarge 4's average refusal rate for malicious attack requests is 96.6%, the highest among open-source models.
50%UnverifiedLarge 4 has image comprehension and ultra-long document reading capabilities.
50%UnverifiedThe open-weight model is free for individuals, schools, and normal business use, but cloud platforms reselling the service need authorization from Mistral.
50%UnverifiedLarge 4's context window reaches one million tokens and uses a sparse attention mechanism that attends to at most 2048 of the most relevant positions each time.
50%UnverifiedOn the Artificial Analysis coding agent leaderboard, Large 4 scores 48%, essentially tied with the top open-source model GLM5.3, but uses 47% fewer tokens per task on average.
50%UnverifiedOn TerminalBench 4.0, Large 4 scores 41%, a 13-percentage-point improvement over the preview version from three weeks earlier, and Mistral says it ranks first among open-source models.
50%UnverifiedHugging Face testing shows that compressing Large 4 into a two-bit version requires an Apple Ultra computer with 512GB of unified memory to run.
50%UnverifiedRunning Large 4 on a typical laptop with 16GB/32GB of memory generates at about 15 tokens per second.
50%UnverifiedFrench AI company Mistral has opened the weights of its one-trillion-parameter Large 4 model for public download.
50%UnverifiedLarge 4 was released in preview earlier this month, and after three more weeks of training, the improved version's weights were published on October 29.
50%UnverifiedMistral released the native multimodal model Large 4, with approximately 1 trillion total parameters; API access is now open, with weights planned for release by the end of the month.
50%