Seed-Coder-8B
8B open code models (base, instruct, reasoning) whose pretraining data was scored and filtered by an LLM instead of hand-written rules.
A model-centric data pipeline replaces per-language heuristics; the reasoning variant adds long-CoT RL. Authors claim SOTA among open models of similar size and wins over some larger ones. MIT licence.
- Date
- Wednesday, 4 June 2025
- Lab
- ByteDance Seed
- Kind
- open-weights
- Access
- open weights
Paper v1 2025-06-04; HF weights appeared 2025-04-29 UTC, about five weeks before the paper.
Sources
This record was checked against its sources on 6 October 2026. How we check