SupraLabs Releases 50M Reasoning Model

SupraLabs has released the Supra-50M-Reasoning (ThinkSupra-50M) model as part of Project Chimera. This model is a reasoning-specialized version of Supra-50M-Instruct, designed to first generate a thinking chain before outputting an answer.

Key technical information:

• Training method: The Supra-50M-Base model was SFT (Supervised Fine-Tuning)-trained for 6 epochs using a synthetic dataset of 500 examples (SupraThink-Dataset-500x) generated with Qwen3 1.7B.
• Response structure: The thinking process is marked with <|begin_of_thought|> and <|end_of_thought|> tags, and the final answer is separated out with a <|begin_of_solution|> tag.
• Specifications: It uses bfloat16 precision, and as it is currently experimental, hallucinations may occur.

SupraLabs plans to sequentially release 124M and 350M-scale Base, Chat, Reasoning, and Coding models in the future.