Nanbeige ships a 33B open-weight model built to teach smaller coding agents
Nanbeige4.3-33B is Apache 2.0 licensed, handles a 512K-token context, and is designed as a teacher model for distilling agentic coding skills into smaller models.
Nanbeige released Nanbeige4.3-33B this week, a 33 billion parameter dense model licensed under Apache 2.0. It supports a context window of up to 524,288 tokens, and ships in BF16 safetensors with a chat template that handles configurable reasoning and tool use.
The model card positions it as “the flagship teacher model for the Nanbeige4.3 series”, meant for distilling agentic coding capability into smaller models rather than running as the end product itself. On Nanbeige’s own benchmarks it scores 80.9 on Terminal-Bench 2.1, 65.5 on DeepSWE, 60.3 on SWE-Bench Pro, and 89.5 on GPQA Diamond.
Why it matters: a permissively licensed 33B model with a genuinely long context window is a reasonable base for teams who want to self-host a coding assistant or build their own smaller fine-tunes without negotiating a licence. The teacher-model framing also tells you what it is for: distillation pipelines, not necessarily a model you’d run as-is in production.
The caveat: every benchmark figure here is Nanbeige’s own. There is no independent evaluation yet to check Terminal-Bench or SWE-Bench Pro scores against.