The Pulse
Nanbeige Releases 33B Open Model With 512K Context
Nanbeige has released Nanbeige4.3-33B, an Apache 2.0 model with a 524,288-token context window. Its model card reports strong scores on agentic coding and workplace-task benchmarks, while describing the model as a teacher for a planned smal
AI.info Team · Written with AI · fact-checked ·
“Our strategy centers on building a high-capacity teacher agent first, which will subsequently anchor the distillation of advanced agentic capabilities into smaller models.”
Nanbeige Team, model card
A 33B model built to train smaller agents
Nanbeige has released Nanbeige4.3-33B, an open-weights language model with 33 billion parameters and support for a context window of up to 524,288 tokens. The company presents it as the flagship teacher model in its Nanbeige4.3 series: a larger system intended to help develop smaller models, including a planned Nanbeige4.3-3B. The model card lists an Apache 2.0 license and BF16 weights.
Nanbeige pitches the release at complex coding and professional “cowork” tasks, where models use tools to complete multi-step work. Its card describes the system as a starting point for distilling agent skills into smaller models, rather than the end of the series. Nanbeige4.3-3B is an example of that future work, not a model included in this release.
How the 512K window fits the serving setup
The stated context limit is 512K tokens, but Nanbeige recommends different generation settings by task. For agent and tool-use work, it suggests a maximum of 65,536 new tokens; for reasoning and chat, it recommends up to 262,144. Those are generation limits in the model card, distinct from the model’s maximum context length.
The release includes instructions for running the weights with Transformers, SGLang and vLLM. Its chat template can turn reasoning on or off and control whether reasoning from earlier assistant turns carries into later turns. Nanbeige recommends retaining that history for multi-turn tool use, office work and coding-agent workflows, and says XML is its preferred format for tool calls, with JSON available for compatibility.
Strong scores, with clear exceptions
Nanbeige compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, saying all four were evaluated at their highest thinking effort with the same inference hyperparameters and harness for each benchmark. The lab reports 80.9 on Terminal-Bench 2.1, ahead of Qwen’s 73.0, and 65.5 on DeepSWE, compared with Qwen’s 42.5. On Automation Bench, Nanbeige lists 54.2 against Qwen’s 38.3.
The table is not a clean sweep. Qwen3.8-27B scores 61.7 on SWE-Bench Pro, above Nanbeige4.3-33B’s 60.3, and 86.4 on GDPval, compared with Nanbeige’s 84.8. Gemma4-31B leads on GPQA Diamond, scoring 91.3 to Nanbeige’s 89.5. The card also notes that benchmark procedures differ: for example, DeepSWE uses mini-swe-agent, while NL2Repo uses Claude Code.
The benchmark table is the lab’s case
The published results make Nanbeige4.3-33B a notable entry among models near the 30-billion-parameter mark, particularly on several coding and agent benchmarks. They remain the company’s own comparison, however; the model card supplies scores and evaluation details, not an independent replication. Results on any one benchmark also do not establish how the model will perform across other software, tools or workplace tasks.
For now, the release combines a long context window and downloadable BF16 weights with a plan to use the 33B model as a teacher. Nanbeige’s stated next step is distillation into smaller systems, with Nanbeige4.3-3B named as an example.