Nanbeige4.3-33B: Benchmarks for the New Open Agentic Coding Model
Nanbeige4.3-33B's benchmark scores versus Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B on agentic coding and cowork tasks, explained.
What is Nanbeige4.3-33B?
Nanbeige4.3-33B is a 33 billion parameter open-weight language model released by the Nanbeige Team, built as the flagship “teacher” model for their Nanbeige4.3 series. Rather than being positioned as a general chatbot, it’s designed specifically for agentic coding and professional cowork tasks, the kind of multi-step, tool-using work that involves writing and debugging code, navigating repositories, and completing office-style jobs end to end. According to its model card, it posts the best results among open models in the roughly 30B parameter class across most of the benchmarks it was tested on.
TL;DR
- Nanbeige4.3-33B is a 33B open-weight model (Apache 2.0 licensed) released on Hugging Face as a “teacher” model meant to anchor future, smaller distilled models like a planned Nanbeige4.3-3B.
- It beats Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B on most of the 17 agentic benchmarks reported in its model card, particularly in terminal and software engineering tasks.
- On Terminal-Bench 4.0 it scores 23.8 versus 6.3 for Qwen3.8-27B and near-zero for the other two, a wide gap on one of the harder agentic coding tests.
- It loses narrowly to Qwen3.8-27B on SWE-Bench Pro (60.3 vs 61.7) and GDPval (84.8 vs 86.4), showing the race at this model size is not a clean sweep.
- The model supports a 512K token context window and ships with configurable chat-template flags for thinking mode and tool-call formatting (XML or JSON).
- It’s deployable through Transformers, SGLang, or vLLM, with documented tensor-parallel server configs for self-hosting.
- As of its release the model had already drawn 784 likes and over 9,600 downloads on Hugging Face, reflecting notable interest from the open-model community.
Remy.
What benchmarks did Nanbeige4.3-33B get tested on?
The model card groups results into three categories: agentic coding, agentic cowork, and general reasoning. Agentic coding benchmarks include Terminal-Bench 2.1 and 4.0, DeepSWE, SWE-Bench Pro, SWE-Atlas-QnA, NL2Repo, Scicode, and SkillsBench. These test whether a model can operate a terminal, navigate and modify real code repositories, and solve software engineering tasks that go beyond single-turn code generation.
Agentic cowork benchmarks look more like office and business work: Automation Bench, JobBench, GDPval, Claw-Eval, Toolathlon-Verified, DeepPlanning, EnterpriseOps-Gym, YC-bench, and Tau3-Bench. These simulate tasks such as planning multi-step workflows, completing professional deliverables, and handling tool-augmented dialogue across many turns. Finally, GPQA Diamond, AIME 2026, and AA-LCR cover more traditional reasoning and math performance, serving as a sanity check that agentic specialization hasn’t come at the cost of general problem-solving.
All four models in the comparison, Nanbeige4.3-33B, Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, were evaluated at their highest thinking effort setting, using the same inference hyperparameters and test harness per benchmark. That consistency matters for comparing agentic models fairly, since different harnesses (mini-swe-agent, Claude Code, Terminus 2) can meaningfully change scores on the same underlying model.
How does it compare to Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B?
Nanbeige4.3-33B leads on the majority of the 17 benchmarks listed. The gaps are largest on terminal-focused tests: Terminal-Bench 4.0 shows Nanbeige4.3-33B at 23.8 against Qwen3.8-27B’s 6.3, Muse-Glimmer-30B’s 1.6, and Gemma4-31B’s 0.0. DeepSWE tells a similar story, 65.5 for Nanbeige4.3-33B versus 42.5 for Qwen3.8-27B and single digits for the other two. These are the kinds of tasks where a model has to actually operate a sandboxed environment correctly over many steps, not just produce a plausible code snippet.
Qwen3.8-27B is the closest competitor overall and actually wins on a couple of the harder software engineering and professional-task benchmarks: SWE-Bench Pro (61.7 vs 60.3) and GDPval (86.4 vs 84.8). Gemma4-31B edges out both on GPQA Diamond (91.3), suggesting it may carry a reasoning edge in some domains, even while trailing significantly on nearly every agentic task. Muse-Glimmer-30B consistently lands in third or fourth place across most categories.
On reasoning benchmarks the four models are closer together than on agentic ones. AIME 2026 scores sit in a tight band from 89.2 to 95.2, and AA-LCR ranges from 72.0 to 84.0. This suggests the differentiation among these ~30B models shows up mainly in how well they execute long, tool-using workflows, not in raw math or knowledge-style reasoning.
What does the YC-bench and JobBench data actually show?
Two benchmarks stand out for measuring something closer to real economic value rather than pass/fail accuracy. YC-bench reports a dollar figure, and Nanbeige4.3-33B’s reported score of $1,025,437 is notably higher than Qwen3.8-27B’s $641,290, Muse-Glimmer-30B’s $513,978, and Gemma4-31B’s $265,152. The model card doesn’t detail exactly what this dollar figure represents beyond the benchmark name, so it should be read as a relative ranking signal rather than a literal cash value, but the gap is large enough to be a meaningful data point for buyers evaluating agentic ROI.
JobBench evaluates performance across 35 different occupations using a Claude Code harness equipped with a search tool. Nanbeige4.3-33B scored 44.8 overall versus 34.5 for Qwen3.8-27B, and the model card notes it outperforms the baseline in most of the 35 occupations tested. That breadth across professions, rather than just coding, is part of what the “cowork” framing is trying to capture: the model is being pitched not just as a coding assistant but as something closer to a general professional agent.
Is Nanbeige4.3-33B worth deploying?
For teams specifically building agentic coding tools or multi-step automation products, Nanbeige4.3-33B’s benchmark profile is compelling at its size class. The jump on Terminal-Bench 4.0 and DeepSWE in particular suggests it handles long-horizon, tool-using tasks more reliably than its closest open competitors around 30B parameters. It’s also licensed under Apache 2.0, which removes licensing friction for commercial deployment, and it supports a 512K token context window, useful for tasks involving large codebases or long multi-turn tool-use sessions.
The tradeoffs are real too. It doesn’t universally win, losing to Qwen3.8-27B on SWE-Bench Pro and GDPval, both of which are demanding, practically-oriented benchmarks. Teams whose workloads map closely to those specific tasks should weigh that against Nanbeige4.3-33B’s wins elsewhere. The model card also recommends different temperature, top_p, and max token settings depending on whether you’re running agentic/tool-use workloads (temperature 1.0, 65,536 max tokens) versus reasoning and chat tasks (temperature 0.6, 262,144 max tokens), so getting inference configuration right matters for reproducing the reported results.
Deployment is straightforward for teams already running open-weight infrastructure: the model supports Transformers, SGLang, and vLLM, with published commands for launching tensor-parallel servers (tp-size 4 in the reference configs) and built-in tool-call parsing compatible with the Qwen3 reasoning and tool-call format.
Frequently Asked Questions
What is Nanbeige4.3-33B designed for?
It’s built primarily for agentic coding and professional “cowork” tasks, meaning multi-step tool use, software engineering, and office-style workflows, rather than general chat. The Nanbeige Team also positions it as a teacher model intended to anchor distillation into smaller models in the future.
How big is Nanbeige4.3-33B and what’s its license?
It has 33 billion parameters and is released under the Apache 2.0 license, based on a llama-architecture base model (Nanbeige4.3-33B-Base) fine-tuned for chat and agentic use.
How does Nanbeige4.3-33B compare to Qwen3.8-27B?
Nanbeige4.3-33B beats Qwen3.8-27B on most of the 17 reported benchmarks, especially terminal and coding agent tasks like Terminal-Bench 4.0 and DeepSWE. Qwen3.8-27B wins on SWE-Bench Pro and GDPval, so results vary by task type.
What context length does Nanbeige4.3-33B support?
Up to 524,288 tokens (512K), according to its model card.
How can I run Nanbeige4.3-33B?
The model card documents deployment through Hugging Face Transformers for basic inference, and through SGLang or vLLM for serving, including tensor-parallel server configurations and compatible reasoning/tool-call parsers.
