LLM Council (documentation: https://github.com/machine-theory/lm-council?tab=readme-ov-file, demonstration: https://huggingface.co/spaces/burtenshaw/karpathy-llm-council and video presentation of related research: <a href="https://www.youtube.com/watch?v=hI0XCE27QqE”>https://www.YouTube. com/watch?v=hI0XCE27QqE is a “democratic” space that synthesizes the responses to a query from several language models into one.

The final generation goes through several steps orchestrated by a primary agent:

• Depending on the instruction, up to 4 models are chosen to produce a response. During testing, GPT Oss, Qwen 3 and Deepseek V3 were implemented.
• Each language model produces its response individually.
• Each LLM evaluates the anonymous responses of the others, a final ranking is established
• A compilation makes it possible to obtain the final synthesis.

Free and open-source.

Test “Generative AI in education, uses and limits”:

Full Generation: https://docs.numerique.gouv.fr/docs/6e3fbb30-1fbe-49d0-9c0f-24d67cd569b1/

Test “Generative AI training 6h”

Full generation: https://docs.numerique.gouv.fr/docs/33a6778c-ddb3-446e-bd6c-6088b2cbc7da/