Mistral AI Drops 1-Trillion-Parameter Large 4 Nicknamed “Le Chonk”
European open-source pride just reached absurd proportions. The Parisians at Mistral AI decided modesty is dead, unleashing a heavyweight monster designed to humble the closed corporate labs of Silicon Valley.
The French AI champion pushed the public preview of Mistral Large 4, officially embracing the community moniker "Le Chonk" across its official developer channels. Built on a sparse Mixture-of-Experts architecture, the system packs 1.05 trillion parameters in total, though only 49 billion activate during any single token pass to keep server farms from melting under pressure. The setup also integrates a 1-million-token context window alongside an internal 1.6-billion-parameter vision encoder.
Skipping rented American cloud platforms, the team trained the model entirely from scratch on 3,800 Nvidia Grace Blackwell GPUs situated inside European datacenters, training across more than 160 languages. This arrival lands roughly ten months after its previous flagship iteration, which carried a comparatively modest 675 billion parameters.
In developer tests, Large 4 scored 61.7% on the DeepSWE v1.1 coding benchmark, overtaking rivals like DeepSeek V4 Pro and Qwen3.8 Max while landing just behind Kimi K3. When shoved into real command-line environments on Terminal-Bench 4.0, the model logged roughly double the performance of its peers, although blind human evaluations run by Surge AI placed its code output in second place with a 3.74 out of 5 score, trailing Anthropic's Claude Opus 5.
Digital defense turned into the model's main flex. On the CyberGym-E2E exploit assessment, Large 4 locked down first place with an 82% score, reproducing and fixing real software vulnerabilities faster than MiMo-V2.6-Pro and Grok 4.7. The preview currently runs in developer APIs at $0.68 per million input tokens and $2.09 per million output tokens, with fully open weights scheduled for public release in late October.
Dropping an open trillion-parameter powerhouse into the wild either marks the ultimate democratic gift to global software engineering or a chaotic cybersecurity gamble that corporate boardrooms were wholly unprepared for. Proprietary software moats suddenly look considerably shallower than tech giants hoped.
Source: Mistral AI
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.