SupraLabs releases Supra-50M-Reasoning with chain-of-thought reasoning
A 50M parameter open model that generates a full thinking chain before every answer.
SupraLabs has released Supra-50M-Reasoning (ThinkSupra-50M), a 50 million parameter language model that produces a step-by-step thinking chain before every answer. Fine-tuned from Supra-50M-Base using supervised fine-tuning (SFT) with bfloat16 precision, the model was trained for 6 epochs on a custom synthetic dataset of just 500 samples generated by Qwen3 1.7B. Each response follows a strict format: a thinking block enclosed in <|begin_of_thought|> and <|end_of_thought|> tags, followed by the final solution inside <|begin_of_solution|> and <|end_of_solution|>.
The model is experimental and the creators openly acknowledge hallucinations, but it demonstrates that chain-of-thought reasoning can be achieved even at the 50M parameter scale using small, synthetic datasets. It is fully open source under the Supra-50M collection (Project Chimera), with inference code provided via the Transformers pipeline. The project also teases upcoming releases: Supra-124M (Base, Chat, Reasoning) and Supra-350M (adding a Coding variant). This release lowers the barrier for experimenting with structured reasoning in tiny models, though production use is not yet recommended due to reliability concerns.
- Fine-tuned from Supra-50M-Base on 500 synthetic reasoning samples from Qwen3 1.7B over 6 epochs using SFT with bfloat16.
- Outputs a formatted chain-of-thought before every answer: <|begin_of_thought|> ... <|end_of_thought|> <|begin_of_solution|> ... <|end_of_solution|>.
- Upcoming model sizes include Supra-124M and Supra-350M with reasoning, chat, and coding variants under Project Chimera.
Why It Matters
Demonstrates that chain-of-thought reasoning can be distilled into tiny models, opening up experimentation for constrained environments.