Mistral Large 4「Le Chonk」開放權重現可供下載
Mistral’s New ‘Le Chonk’ AI Model Is Big, Open and Now Available for Download
Mistral AI 已開放旗艦模型 Mistral Large 4(ML4,暱稱「Le Chonk」)下載;該模型此前於本月初進入公開預覽。Mistral 表示,模型在 SWE-bench 和 Terminal Bench 等 Agent 基準達到「state-of-the-art」表現,並具備網絡安全、編碼及製造業和金融任務方面的能力。
New AI models are coming, and they don’t stop coming.
On the heels of the recent rollouts of Meta’s Muse and Google’s Gemini 4 Carbon, Mistral AI is launching its new flagship frontier model, Mistral Large 4. The model – affectionately known as “Le Chonk” – was available in public preview earlier this month. Now, after weeks of additional reinforcement learning, fine-tuning and assessing it with feedback from developers, cybersecurity experts and state officials, the Mistral team announced Thursday that ML4 is available for download.
“The response to the preview has exceeded our expectations,” Mistral said in its announcement. “We’ve put that feedback to work while continuing to scale reinforcement learning, driving massive improvements in coding and defensive cybersecurity.”
More from CNET
A French company known for its open-source and open-weight LLMs, Mistral AI reports that the Large 4 model “ranks among the best open-weight models of the world” on aggregated metrics and also outperforms European and US models, while also catching up to Chinese competitors, who have been dominating the open-weight sector.
Earlier this month, I spoke with Mistral Chief Scientist Guillaume Lample and VP of Science Pierre Stock about the new model – what they’re calling the best open model built in the West – and its specialized capabilities and extreme compute efficiency. Here’s what to know.
Key capabilities of Mistral Large 4
Lample and Stock said that Mistral Large 4 was built to process both text and images, closing performance gaps from previous Mistral models. Specifically, Large 4 touts more advanced cybersecurity and agentic capabilities, as well as a better ability to handle manufacturing- and finance-related tasks.
Cybersecurity
Mistral Large 4 was designed to excel at defensive cybersecurity tasks, including log inspection, code auditing and checking for newly published exploits or vulnerabilities. Beyond defensive measures, Mistral says the new model offers offensive cybersecurity measures to identify and address weaknesses.
Agentic AI
“Really, the core competency we’re looking at for ML4 is agentic capabilities,” Stock said.
Mistral Large 4 was built with a focus on agentic performance for planning, research or handling complex enterprise tasks that would take human experts hours or days to complete. The Mistral AI team said the model reaches “state-of-the-art” performance on critical agentic benchmarks like SWE-bench and Terminal Bench – evaluation frameworks for testing how well agents handle complex tasks – for tasks like coding, analyzing and pulling requests.
Geospatial and grounding capabilities
Mistral says the model outperforms competitors in aerial and satellite imagery analysis, with the capacity to identify visual patterns and objects on maps, such as forest fires.
“On grounding capabilities, it outperforms all existing models, including the closed ones,” Mistral said when it released the model preview. “These are critical as cyberattacks get faster and cheaper, companies write more code than ever and the world needs more chips to keep up with AI demand.”
Grounding is a technique used in AI development to try to eliminate hallucinations – when the LLM generates inaccurate information – by anchoring the AI’s outputs in external resources, like documents and databases, rather than just the materials it was trained on internally. Grounding also helps keep data and information fresh while providing context for the AI’s responses.
Mistral says the new model was also trained on more than 160 languages, including every official language of the European Union, as well as Latin and non-Latin languages, to answer prompts natively.
If you’re interested in getting started with ML4, it’s available to try and customize in Mistral Studio, through the Mistral API, or on OpenRouter. Mistral also says you can download the open weights from Hugging Face, try ML4 in Vibe Code or access it through Nvidia-hosted endpoints, with additional options to customize for enterprise users.
來源:CNET · cnet.com