Nanbeige4.3-33B: A New Open Agentic Coding Model Benchmarked
Nanbeige4.3-33B is a 33B open-weight model claiming top agentic coding and cowork scores against Qwen3.8-27B and Gemma4-31B.
What is Nanbeige4.3-33B?
Nanbeige4.3-33B is a 33-billion-parameter open-weight language model released by the Nanbeige team under an Apache 2.0 license, built specifically for agentic coding and “cowork” tasks rather than general chat. According to the model’s own benchmark tables, it outperforms similarly sized open models like Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B on most agentic coding and tool-use evaluations, while landing close to parity on general reasoning tests like GPQA Diamond and AIME.
TL;DR
- Nanbeige4.3-33B is positioned as a teacher model, meaning its main purpose is to anchor a distillation pipeline for smaller models, including a planned Nanbeige4.3-3B.
- The model reports leading scores on Terminal-Bench 2.1 and 4.0, DeepSWE, SWE-Atlas-QnA, and several cowork benchmarks, beating Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B on most metrics.
- It does not win everywhere: Qwen3.8-27B edges it out on SWE-Bench Pro and GDPval, and Gemma4-31B leads on GPQA Diamond, showing the gains are concentrated in agentic workflows rather than pure reasoning.
- The model supports a 512K token context window and ships with configurable chat-template flags for thinking mode and tool-call formatting.
- It’s built on a Llama-architecture base (Nanbeige4.3-33B-Base) and is deployable through Transformers, SGLang, or vLLM, with documented tensor-parallel serving configs.
- As of its release, the model had drawn 784 likes and over 9,600 downloads on Hugging Face, a notable amount of attention for a mid-sized open model aimed at a narrow agentic niche.
What does “agentic coding and cowork” actually mean here?
Nanbeige draws a line between two categories of evaluation in its benchmark suite. “Agentic coding” covers tasks where the model has to operate inside a terminal or repository, write and modify code, and complete multi-step software engineering work. This includes Terminal-Bench 2.1 and 4.0, DeepSWE, SWE-Bench Pro, SWE-Atlas-QnA, NL2Repo, Scicode, and SkillsBench. These benchmarks typically wrap the model in an agent harness (the card references tools like Terminus 2, mini-swe-agent, and Claude Code) that lets it issue shell commands, read file output, and iterate.
“Agentic cowork” is a broader category meant to simulate professional, non-coding knowledge work: things like Automation Bench, JobBench, GDPval, Claw-Eval, Toolathlon-Verified, DeepPlanning, EnterpriseOps-Gym, YC-bench, and Tau3-Bench. JobBench in particular evaluates the model across 35 different occupations, and Nanbeige reports that its model beats the baseline in most of them. The inclusion of a benchmark called YC-bench, scored in dollar terms rather than percentages, signals that these evals are trying to approximate business or economic value generated by an agent, not just task completion rates.
How do the benchmark numbers actually compare?
On the coding side, the gaps reported are large in places. Nanbeige4.3-33B scores 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B, and 47.2 for Gemma4-31B. The gap widens dramatically on Terminal-Bench 4.0, where Nanbeige4.3-33B posts 23.8 against Qwen3.8-27B’s 6.3, Muse-Glimmer-30B’s 1.6, and Gemma4-31B’s 0.0, suggesting the newer, harder version of that benchmark separates models much more sharply. DeepSWE shows a similar pattern: 65.5 for Nanbeige4.3-33B versus 42.5 for Qwen3.8-27B and single digits for the other two.
Not every metric favors Nanbeige. On SWE-Bench Pro, Qwen3.8-27B comes out slightly ahead at 61.7 versus Nanbeige’s 60.3. On GDPval, a cowork benchmark, Qwen3.8-27B again leads, 86.4 to 84.8. These are close results, but they matter for anyone deciding between models based on a specific workload rather than an aggregate score.
On pure reasoning benchmarks, the picture is tighter and less one-sided. Gemma4-31B actually posts the top GPQA Diamond score (91.3) despite trailing badly on agentic tasks, which underlines that reasoning ability and agentic tool-use competence are not the same skill and don’t always scale together. Nanbeige4.3-33B leads AIME 2026 (95.2) and AA-LCR (84.0) by modest margins.
All of these comparisons come from Nanbeige’s own model card, evaluated with what it describes as consistent inference hyperparameters and harnesses across models. As with any vendor-published benchmark table, independent reproduction is the real test; the released weights make that possible for anyone with the hardware to run a 33B model.
What’s under the hood?
Nanbeige4.3-33B is built on a Llama-architecture base model called Nanbeige4.3-33B-Base, and the full release is split across 14 safetensors shards. It’s tagged for both English and Chinese, and the repository includes evaluation result files alongside the standard tokenizer and config assets.
The chat template exposes two notable toggles. enable_thinking turns the model’s reasoning trace on or off per response, and it’s on by default. preserve_thinking controls whether that reasoning carries over across turns in a multi-turn conversation. Nanbeige recommends turning preserve_thinking off for ordinary chat and question answering, but on for multi-turn tool use, office-style tasks, and coding-agent workflows, which is also how the benchmark numbers themselves were generated.
For tool calling, the model supports both XML and JSON formats, with XML recommended for best performance. Context length tops out at 524,288 tokens (512K), which is large enough to hold substantial codebases or long multi-turn agent transcripts without truncation.
Is Nanbeige4.3-33B worth deploying?
For teams building coding agents or automation pipelines who want an open-weight alternative to API-only models, the benchmark profile suggests Nanbeige4.3-33B is a reasonable candidate to evaluate, particularly for terminal-based coding agents and tool-calling workflows where its reported scores are furthest ahead of the competition. The Apache 2.0 license removes usage restrictions that come with more closed releases, and the documented recipes for SGLang and vLLM (including tensor-parallel serving across 4 GPUs) make it straightforward to stand up for inference.
That said, a few caveats apply. First, Nanbeige frames this release explicitly as a teacher model for an upcoming distillation pipeline, not necessarily the end product most teams will run in production, since a 33B dense model still carries meaningful VRAM and latency costs compared to smaller distilled variants expected later, like the planned Nanbeige4.3-3B. Second, the benchmark wins aren’t universal; SWE-Bench Pro and GDPval results favor Qwen3.8-27B, so workload-specific testing still matters before committing. Third, all comparative numbers come from the vendor’s own card, and benchmark suites like Automation Bench, JobBench, and YC-bench are less standardized across the industry than something like SWE-Bench, so their comparability to other published leaderboards is harder to verify independently.
Recommended inference settings
Nanbeige publishes separate recommended settings depending on use case. For agentic and tool-use tasks, it suggests a temperature of 1.0, top_p of 0.95, and up to 65,536 new tokens. For reasoning and chat tasks, it recommends a lower temperature of 0.6, the same top_p of 0.95, but a much larger max new tokens budget of 262,144, reflecting the longer chains of thought those tasks can require.
Frequently Asked Questions
What is Nanbeige4.3-33B used for?
It’s an open-weight language model optimized for agentic coding (operating in terminals and repositories to write and debug code) and “cowork” tasks that simulate professional knowledge work across different occupations.
How does Nanbeige4.3-33B compare to Qwen3.8-27B?
Nanbeige4.3-33B leads on most agentic coding benchmarks reported on its model card, including Terminal-Bench 2.1, Terminal-Bench 4.0, and DeepSWE, but Qwen3.8-27B scores higher on SWE-Bench Pro and GDPval. Reasoning benchmark results are close between the two.
Is Nanbeige4.3-33B open source?
Yes. It’s released under an Apache 2.0 license on Hugging Face, with full safetensors weights, a base model (Nanbeige4.3-33B-Base), and documented deployment instructions for Transformers, SGLang, and vLLM.
What context length does Nanbeige4.3-33B support?
Up to 524,288 tokens (512K), according to the model card.
Why does Nanbeige call this a “teacher model”?
The team describes Nanbeige4.3-33B as the flagship model in a series meant to anchor a distillation pipeline, where its agentic capabilities will be distilled into smaller models, with a 3B variant mentioned as a future release.
