Le Chonk
A model for visual grounding tasks, focused on object localization in dense scenes. It achieves 42% on the Dense 200 benchmark and is claimed to surpass frontier closed-source models.
Timeline (last 90 days)
Mistral AI released its latest-generation flagship large model Large 4 in Open Weight form, under the internal codename "Le Chonk".
Mistral's Large 4 weights (codename 'Le Chonk', 1T total parameters) will be released before the end of the month.
Le Chonk achieved a score of 42% on the Dense 200 benchmark.
Le Chonk is claimed to surpass frontier closed-source models.
When the contact lens case task was re-run using OpenCode's "no reasoning" mode, Le Chonk successfully produced a box whose left and right lids can open independently.
Le Chonk scored a total of 42/80, or 52.5%, equivalent to 5.25/10, across the eight questions in KingBench3.
The tester recommends Le Chonk for controlled visualization prototyping with confidence when paired with a browser and testing loop, but does not recommend it as a default coding model for longer, less error-tolerant work.
When browser error messages were passed back to Le Chonk verbatim, it successfully fixed the module path issue in the 3D wristwatch project, scoring 6/10 after the fix.
Le Chonk scored 9.5/10 on the elevator simulation task, delivering a functional three-elevator setup with genuinely working queue logic.
Le Chonk scored 10/10 on the local fine-tuning task, completing the full pipeline of generating a panda knowledge dataset, locally fine-tuning a Gemma 2B model, training a LoRA adapter for 400 steps, and reducing validation loss from approximately 7.97 to 0.31.
1 more timeline events
All Facts (9)
Le Chonk scored a total of 42/80, or 52.5%, equivalent to 5.25/10, across the eight questions in KingBench3.
50%UnverifiedLe Chonk scored 9.5/10 on the elevator simulation task, delivering a functional three-elevator setup with genuinely working queue logic.
50%UnverifiedLe Chonk scored 10/10 on the local fine-tuning task, completing the full pipeline of generating a panda knowledge dataset, locally fine-tuning a Gemma 2B model, training a LoRA adapter for 400 steps, and reducing validation loss from approximately 7.97 to 0.31.
50%UnverifiedIn the default OpenCode run, Le Chonk used up the entire reasoning round on the contact lens case task without creating any files, scoring zero points.
50%UnverifiedWhen the contact lens case task was re-run using OpenCode's "no reasoning" mode, Le Chonk successfully produced a box whose left and right lids can open independently.
50%UnverifiedWhen browser error messages were passed back to Le Chonk verbatim, it successfully fixed the module path issue in the 3D wristwatch project, scoring 6/10 after the fix.
50%UnverifiedThe tester recommends Le Chonk for controlled visualization prototyping with confidence when paired with a browser and testing loop, but does not recommend it as a default coding model for longer, less error-tolerant work.
50%UnverifiedMistral AI released its latest-generation flagship large model Large 4 in Open Weight form, under the internal codename "Le Chonk".
50%UnverifiedMistral's Large 4 weights (codename 'Le Chonk', 1T total parameters) will be released before the end of the month.
50%