Nanbeige4.3-33B Launches with Open Weights, Aims to Distill Agent Capabilities into a Smaller Model

The Nanbeige Large Model Lab under BOSS Zhipin has released Nanbeige4.3‑33B, a dense language model with approximately 33 billion parameters, supporting up to 512K token context length. It is specifically optimized for programming agents, tool usage, and professional office tasks, and is being made publicly available under the Apache 2.0 license, including both the conversational and base models. Official evaluations show it ranked first in 17 out of 20 benchmarks comparing models of around 30B scale—for example, scoring 80.9 on Terminal‑Bench 2.1 and 65.5 on DeepSWE—but still lags behind some competitors on SWE‑Bench Pro, GDPval, and GPQA Diamond. Of greater interest is the release strategy: Nanbeige first trained a strong 33B “teacher model,” then distilled its agent capabilities into the upcoming Nanbeige4.3‑3B. However, all current results were generated by the releasing team and have not yet been independently replicated; the model card also does not disclose training data, computational resources, reinforcement learning methods, or comprehensive safety evaluations.

Article author and source: Nanbeige Team

Nanbeige4.3 follows a different approach from the common method of directly training small models: it first constructs a larger teacher model with stronger agent capabilities, then uses the data, trajectories, or other supervision signals it generates to train smaller models.

The Nanbeige 4.3-33B released this time is this teacher. The team says the subsequent distillation pipeline will start from Nanbeige 4.3-3B, aiming to compress programming, tool usage, and long-task execution capabilities into a more deployable size. The 3B model has not yet been released, and the release date, scale of distillation data, and how much of the teacher’s capabilities will be retained have not been disclosed.Hugging Face

The practical significance of this approach lies in the fact that, although the 33-billion-parameter model is much smaller than flagship models with hundreds of billions of parameters, it is still not lightweight. Nanbeige4.3‑33B is a dense model, meaning all parameters are used during each token generation, unlike mixture-of-experts models that activate only a subset of experts.

The model weights are in BF16 format. Roughly estimating 2 bytes per parameter, the weights alone require approximately 66 GB of storage or VRAM; actual inference requires additional capacity for KV Cache and intermediate states. Official examples from SGLang and vLLM both use four-GPU tensor parallelism, indicating that a "30B-class open model" does not equate to unconditional full-context execution on standard consumer GPUs; quantization can lower the barrier, but speed, accuracy, and 512K long-context capability must be verified separately.

Nanbeige compares the new model with Qwen3.8‑27B, Muse‑Glimmer‑30B, and Gemma4‑31B. The official statement says all models were tested with maximum thought intensity enabled, under the same set of inference parameters and Agent framework.

In programming agents, Nanbeige4.3‑33B scores 80.9 on Terminal‑Bench 2.1, while Qwen3.8‑27B scores 73.0; on Terminal‑Bench 4.0, the scores are 23.8 and 6.3 respectively; on DeepSWE, they are 65.5 and 42.5. Nanbeige also leads in projects such as NL2Repo, Scicode, and SkillsBench.

However, the same table also retains unfavorable results: on SWE-Bench Pro, Nanbeige scored 60.3, below Qwen’s 61.7; on GDPval, it scored 84.8, below Qwen’s 86.4; and in the scientific reasoning test GPQA Diamond, Nanbeige scored 89.5, while Gemma4-31B scored 91.3.

Therefore, the official claim of “state-of-the-art performance among approximately 30B open models” can only be understood as a comprehensive conclusion under specified opponents, specified agent frameworks, and specified inference settings; it cannot be extended to mean “outperforming all open models,” nor does it demonstrate superiority over proprietary flagship models.

The evaluation environment itself also affects the results. For example, Terminal-Bench 2.1 uses Terminus 2; Terminal-Bench 4.0 and DeepSWE use mini-swe-agent; NL2Repo runs on the Claude Code framework; and SWE-Atlas-QnA is scored by GPT-6 Sol according to 1,304 task-specific rules. Model capabilities, agent scaffolding, tool permissions, and the effectiveness of the grading model cannot be fully separated.

These results have been published solely by the Nanbeige team, and no third party has yet completed a full reproduction under the same hardware, token budget, and scaffolding.

Nanbeige refers to another set of tests as Agentic Cowork, covering office automation, search, planning, enterprise operations, and multi-turn tool calls.

Among these, the Automation Bench scores are 54.2 for KuCoin and 38.3 for Qwen3.8‑27B; JobBench scores are 44.8 versus 34.5; DeepPlanning scores are 63.3 versus 49.2. The lead on Tau3‑Bench and EnterpriseOps‑Gym is minimal, at just 0.3 and 0.8 points respectively.

JobBench now covers 35 occupational categories. The publisher states that the model leads in most occupational categories, particularly in technical writing, website management, content creation, and online business tasks; however, it still lags significantly behind Qwen in categories such as healthcare administration, administrative assistance, and customer service. In other words, a higher overall occupational task score does not mean it can equally reliably perform every type of office work.

The model supports a maximum context of 524,288 tokens. Its chat template provides two switches related to Agent:

enable_thinkingDetermines whether to generate reasoning content for the current response;preserve_thinkingDetermines whether to retain previous reasoning records in multi-turn conversations. The official recommendation is to disable the latter for general Q&A, and enable it for continuous tool usage, office, and code Agent workflows. Additionally, the model supports both XML and JSON tool call formats, with XML being the officially recommended option.

Retaining prior reasoning can reduce the issue of repeatedly rebuilding plans in long tasks, but it also expands the context, increases reasoning costs, and may carry early errors into subsequent steps. 512K is the maximum capacity claimed by the model, but this does not mean it maintains the same recall rate across all positions, nor does it imply that using the full window on standard hardware remains economical.

The official claims the model offers good cost efficiency among mainstream Agent frameworks, but neither the original post nor the model card has published actual throughput, token consumption per task, hardware configuration comparisons, or API pricing. Therefore, "cost efficient" cannot currently be quantitatively verified like benchmark results.

Both the dialogue model and the base model are licensed under the Apache 2.0 license. Compared to open-weight models that permit only research and prohibit commercial use, this license generally allows commercial use, modification, and redistribution, provided that users comply with specific conditions such as copyright notices, the license text, and trademarks. Weights in Safetensors format are provided for both models, along with instructions for running them using Transformers, SGLang, and vLLM.

The Base version enables other teams to perform instruction fine-tuning, domain-specific post-training, or agent reinforcement learning on their own. According to its own evaluation, it outperforms Gemma4-31B-Base across eight foundational capability benchmarks, including MMLU-Pro, GPQA, Mathematics, MBPP, MMLU, and CMMLU. However, since only this single model was used for comparison, it is insufficient to establish a comprehensive ranking of foundational models.

The model card currently does not disclose the composition of the pretraining data, the data cutoff date, the number of training tokens, the training compute, the source of post-training trajectories, or the safety testing performed. It also does not publish a complete technical report. As a result, external parties can examine the weights, deployment methods, and output behaviors, but cannot determine whether performance improvements primarily stem from pretraining scale, data quality, reinforcement learning, tool trajectories, or specific evaluation optimizations.

What makes Nanbeige4.3‑33B truly noteworthy is not merely topping a single leaderboard, but clearly establishing “Large Teacher Agent—Small Model Distillation” as a defined product roadmap. This strategy will be critically validated only if subsequent 3B models can retain most tool-use and long-task capabilities at significantly lower hardware costs; until then, the 33B model serves as a downloadable, commercially viable, yet independently verifiable teacher starting point for the community.