Nanbeige4.3-33B
AvailableNanbeige4.3-33B is an open-weight agentic coding / reasoning model from Chinese lab Nanbeige, posted on Hugging Face 2026-10-24 together with Nanbeige4.3-33B-Base. Dense 33.1B parameters (Llama architecture, 52 layers, 56 attention heads, 8 KV heads), 524,288-token context, reasoning on by default, tool use; English and Chinese. Apache-2.0. ~66 GB in BF16, ~20 GB at 4-bit. Nanbeige positions it as a flagship 'teacher' to distill into a smaller Nanbeige4.3-3B. Lab-reported (max thinking) vs Qwen3.8-27B: wins 18 of 20 benchmarks, e.g. DeepSWE 65.5 vs 42.5, with leads on Terminal-Bench 4.0 and AutomationBench; trails on SWE-Bench Pro (60.3 vs 61.7) and GDPval (84.8 vs 86.4). Also compared against Muse-Glimmer-30B and Gemma4-31B. Not independently verified.
Specifications
- License
- Open source · Apache-2.0
- Weights
- Downloadable
- Architecture
- Dense
- Parameters
- 33.1B
- Context window
- 524K tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- —
- Modalities
- TextCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.