Syb (@DegenOfWallSt) on X

What it claims

Nanbeige4.3-33B is a 33.1B dense, Llama-style model from Nanbeige under Apache-2.0, with no revenue cap on commercial use. Nanbeige built it as the teacher for a smaller Nanbeige4.3-3B, which it says is coming but has not released. A base model is released alongside the main weights on Hugging Face.

Evidence

All figures come from Nanbeige's own table, run at the highest thinking effort in the same harness for every model. No independent evaluation exists.

Against Qwen3.8-27B, Nanbeige4.3-33B leads on Terminal-Bench 2.1 (80.9 vs 73.0), Terminal-Bench 4.0 (23.8 vs 6.3), DeepSWE (65.5 vs 42.5) and Automation Bench (54.2 vs 38.3). The Terminal-Bench 4.0 gap is the widest in relative terms, though both scores are low.

It trails Qwen3.8-27B on SWE-Bench Pro (60.3 vs 61.7) and GDPval (84.8 vs 86.4). Both gaps are under two points. On GPQA Diamond the comparison model is Gemma4-31B, not Qwen: Nanbeige scores 89.5 against Gemma4-31B's 91.3. The table therefore does not show a single opponent across all rows.

How to use it

• Weights: BF16 weights are about 66 GB. Nanbeige's SGLang and vLLM examples use tensor parallelism across 4 GPUs.
• Context and thinking: The context window is 512K tokens. Thinking is on by default, and passing enable_thinking=False turns it off.
• Tool calls: Nanbeige recommends XML-format tool calls. JSON also works.
• Transformers: Loading needs trust_remote_code.
• Quantised builds: Community GGUF, MLX and FP8 quants are already on Hugging Face. None of them come from Nanbeige.
• Hosting: No hosted inference provider serves it on Hugging Face yet.

Limits

Every benchmark number is self-reported by Nanbeige, and none has been reproduced. Scores are for the highest thinking effort only. The page gives no figures with thinking disabled or for quantised weights.

At full precision the model does not fit on a single consumer GPU or a laptop. Lower-memory use depends on third-party quants whose quality the page does not measure. Remote-code loading means running Nanbeige's model code locally, which needs review before deployment. The 3B student model is announced, not released.