Startup Fortune reported on October 25, 2026, that Nanbeige had released Nanbeige4.3-33B on Hugging Face during that week. Nanbeige describes the open-weight model as a teacher model for its Nanbeige4.3 series, aimed at coding agents and professional cowork, and lists a maximum context of 524,288 tokens.
What Nanbeige says the model is for
Nanbeige positions Nanbeige4.3-33B as the flagship teacher model in its series, intended to help distill agent capabilities into smaller models. The company gives Nanbeige4.3-3B as a planned example, not a released model. Its model card also describes the model’s intended uses and configuration options.
What its benchmark table reports
Nanbeige’s table reports higher scores for Nanbeige4.3-33B than for Qwen3.8-27B in 18 of 20 benchmark rows. Qwen3.8-27B scores higher on SWE-Bench Pro and GDPval, so the results vary by task.
| Benchmark | Nanbeige4.3-33B | Qwen3.8-27B |
| DeepSWE | 65.5 | 42.5 |
| Terminal-Bench 2.1 | 80.9 | 73.0 |
| SWE-Bench Pro | 60.3 | 61.7 |
| GDPval | 84.8 | 86.4 |
These are scores Nanbeige reports from its own evaluations. The company says the models were evaluated at their highest thinking effort using the same inference hyperparameters and harness for each benchmark. Individual tests also use different setups: for example, Nanbeige lists Terminus 2 for Terminal-Bench 2.1 and mini-swe-agent v2.4.6 for DeepSWE. Its broader table includes evaluations with defined task subsets and scoring rules, such as the 157-task General subset for Claw-Eval and revised rubrics for GDPval.
Settings for agents and chat
Nanbeige recommends different generation settings for agent and tool workflows than for general chat and question answering. Its enable_thinking option is on by default; preserve_thinking controls whether prior assistant reasoning carries into later turns. Nanbeige recommends preserving it for multi-turn tool use and code-agent work, and turning preservation off for general chat and QA.
| Use case | preserve_thinking | temperature | top_p | max_new_tokens |
| Agent and tool use | True | 1.0 | 0.95 | 65,536 |
| Reasoning, chat and QA | False | 0.6 | 0.95 | 262,144 |
For tool calls, Nanbeige recommends XML formatting for its best performance and also supports JSON. These are the model card’s recommendations, not a requirement to use one format in every task.
Local inference options
Nanbeige provides quickstarts for Transformers, SGLang and vLLM. Its Transformers example uses automatic device mapping; the SGLang and vLLM examples set tensor parallelism to four.
SimplyExplain’s English-language video also discusses benchmark claims and local setup options including llama.cpp, LM Studio and Ollama. Its discussion of 24GB GPU memory and context capacity uses arithmetic estimates rather than measurements.
