Nanbeige4.3-33B released: 33B open model with 512K context targets agentic coding and cowork tasks

Nanbeige has released Nanbeige4.3-33B, a 33B-parameter model with a context length of up to 524,288 tokens (512K), published on Hugging Face as the flagship teacher model of the Nanbeige4.3 series. Nanbeige claims it delivers state-of-the-art agentic performance among open models in the ~30B parameter class, particularly on complex coding and professional cowork tasks.

Role in the Nanbeige4.3 series

Nanbeige says its strategy is to build a high-capacity teacher agent first, then use it to distill advanced agentic capabilities into smaller models. Nanbeige4.3-33B is that teacher. The team says it will power an upcoming small-model distillation pipeline, for example for Nanbeige4.3-3B. The model tree lists Nanbeige4.3-33B-Base as the base model, with 2 finetunes and 7 quantizations already listed. Weights are stored in safetensors with BF16 tensors.

Benchmark claims

Nanbeige compared the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. All models were evaluated at their highest thinking effort, using the same inference hyperparameters and harness for each benchmark, according to Nanbeige. These are self-reported results. Nanbeige4.3-33B was run with preserve_thinking=true in the chat template.

Nanbeige states the model leads on most of the benchmarks it ran. The categories were:

• Agentic coding: Terminal-Bench 2.1 (via Terminus 2), Terminal-Bench 4.0 and DeepSWE (both via mini-swe-agent v2.4.6), SWE-Bench Pro, SWE-Atlas-QnA, NL2Repo (via Claude Code), Scicode and SkillsBench.
• Agentic cowork: Automation Bench, JobBench, GDPval (scored with revised rubrics), Claw-Eval, Toolathlon-Verified, DeepPlanning, EnterpriseOps-Gym, YC-bench and Tau3-Bench.
• Reasoning: GPQA Diamond, AIME 2026 and AA-LCR.

The table is not a clean sweep. Qwen3.8-27B scores higher on SWE-Bench Pro and GDPval, and Gemma4-31B scores higher on GPQA Diamond. On YC-bench, Nanbeige4.3-33B scores $1,025,437 against $641,290 for Qwen3.8-27B. Nanbeige notes that YC-bench keeps only the most recent 10 turns of user dialogue as context before each model call.

Several evaluation details matter for comparison:

• SWE-Atlas-QnA answers are graded against 1,304 task-specific rubrics, with GPT-6 Sol as judge at medium reasoning effort.
• Automation Bench v1.0.6 covers 600 publicly released tasks.
• Claw-Eval uses only the General subset, which has 157 single-turn user interaction tasks.
• JobBench uses the Claude Code harness with a search tool. Nanbeige says the model outperforms the baseline in most of the 35 occupations evaluated.
• Toolathlon-Verified sets no max_tokens limit on individual requests.

Usage and deployment

The chat template has two controls. enable_thinking is on by default and can be set to False for non-thinking mode. preserve_thinking retains reasoning from earlier assistant turns. Nanbeige recommends False for general chat and question answering, and True for multi-turn tool use, office tasks and code-agent workflows. For tool calling, it recommends tool_call_format="xml" for best performance, with json supported for compatibility.

Recommended sampling settings:

Scenario	Temperature	Top_p	Max new tokens
Agentic and tool-use tasks	1.0	0.95	65,536
Reasoning and chat tasks	0.6	0.95	262,144

The model runs with Transformers, SGLang or vLLM. The published SGLang and vLLM commands use a tensor-parallel size of 4, the qwen3 reasoning parser and the qwen3_coder tool-call parser. API pricing is not yet disclosed, and no Inference Provider currently deploys the model on Hugging Face.

Nanbeige invites collaborations on domain-specific and industry-focused agents, including joint work on data and training, via nanbeige@kanzhun.com.

What this means

Nanbeige4.3-33B is positioned as infrastructure for the series as well as a standalone model: its stated job is to teach smaller Nanbeige4.3 models, starting with a planned 3B. For teams that want to self-host, the 33B size, BF16 weights and published 4-way tensor-parallel configurations for SGLang and vLLM define the hardware starting point. The 524,288-token context length and the recommended 65,536 max new tokens for agentic work fit long-running code-agent sessions.

The benchmark evidence is vendor-run, and the harness choices (Terminus 2, mini-swe-agent, Claude Code, an LLM judge for SWE-Atlas-QnA) affect how results transfer to other setups. The comparison also shows mixed results: Qwen3.8-27B leads on SWE-Bench Pro and GDPval, and Gemma4-31B leads on GPQA Diamond. Teams should test on their own workloads, with preserve_thinking set as recommended for multi-turn agent use.