Nanbeige4.3-33B scores 80.9 on Terminal-Bench

What changed

Nanbeige, the LLM lab inside recruitment platform BOSS Zhipin, released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license. The model is a 33-billion-parameter dense language model designed for agentic coding and office-oriented collaborative tasks. A base variant was also published for custom post-training. The lab has indicated plans to distill agentic skills into smaller models, beginning with a 3B variant.

The defining capability is context length. Nanbeige4.3-33B supports a 524,288-token window (512K), enabling it to process long tool-call sequences and multi-turn workflows without frequent context resets. For teams deploying inference on dedicated hardware who need an open alternative to proprietary agentic systems, this combination of size and context window is the target use case.

According to Nanbeige’s own benchmarks, the model leads 17 of 20 rows against three near-30B competitors: Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. On Terminal-Bench 2.1, a test for coding agent behavior, it scores 80.9 compared to 73.0 for Qwen3.8-27B. On DeepSWE, which measures end-to-end software engineering task completion, Nanbeige4.3-33B reaches 65.5 against Qwen’s 42.5. The model trails on SWE-Bench Pro (60.3 versus Qwen’s 61.7) and scores 23.8 on Terminal-Bench 4.0.

Architecture follows standard patterns. The model uses 52 transformer layers with grouped-query attention (56 query heads and 8 key-value heads), a vocabulary of 166,912 tokens, and runs all 33 billion parameters on every token without mixture-of-experts routing. The chat template exposes three controls: enable_thinking toggles internal reasoning (on by default), preserve_thinking maintains reasoning across turns for multi-step workflows, and tool_call enables XML-formatted function calling.

Why it matters

Dense 30B models have become a reference point for open-source agent systems. Proprietary alternatives like Claude and GPT-class models dominate production agentic workflows but carry API costs, per-token fees, and data residency constraints. An open 33B baseline with 512K context removes friction for teams building internal tools or deploying agents on private infrastructure.

Nanbeige4.3-33B addresses two specific user groups. First, teams deploying multi-turn tool-calling agents who require full conversation history in context. The 512K window accommodates long sequences of API calls, results, and reasoning without windowing. Second, companies required to run models on private infrastructure or unwilling to accept per-token API costs at inference scale.

The benchmark results, particularly on Terminal-Bench 2.1 and DeepSWE, suggest optimization for agentic behaviors: tool calling, multi-step reasoning, and recovery from tool errors. This differs from general-purpose benchmarks where reasoning-focused models excel but may struggle with the repeated stop-and-restart patterns of real agent loops.

The release also signals a longer-term strategy. By planning distillation into smaller models, Nanbeige mirrors approaches used by other labs but on the open-source side. If distillation preserves agentic skills, smaller variants could run on consumer GPUs or edge devices, opening new deployment scenarios. However, the 3B variant remains unreleased, and no timeline appears in the source material.

What to test

Before adopting Nanbeige4.3-33B, verify several vendor claims and practical constraints:

• Benchmark independence. Nanbeige’s model card provides results it measured directly. However, some reported figures for competitor models differ from those competitors’ published results. For example, Nanbeige lists Muse-Glimmer-30B at 61.8 on Terminal-Bench 2.1, while Meta reports 51.7. The SWE-Bench Pro numbers appear potentially swapped between Meta’s Muse-Glimmer-30B and Gemma4-31B figures. Run independent benchmarks using the same harness and hyperparameters before final model selection.

• Memory and latency in practice. BF16 weights require approximately 66 GB. At full 512K context, the KV cache requires roughly 112 GB in BF16 precision. Verify that your hardware stack and quantization strategy meet latency targets. Deployment examples use 4-way tensor parallelism across multiple GPUs.

• Reasoning token overhead. The model defaults to enable_thinking=true, which generates internal reasoning tokens. Measure the impact on token throughput and inference cost in your environment. Reasoning often adds 20-40 percent to latency depending on the implementation.

• Tool-call error handling. The XML tool-call format is recommended. Test how the model behaves when tool calls are malformed or tools return errors. Agent performance often depends on graceful failure handling rather than initial accuracy alone.

• Distillation timeline. The 3B variant is planned but not yet released. If your target deployment is edge or CPU-constrained, confirm release timelines with the lab and whether distillation will preserve agentic task performance.

• Benchmark harness differences. Some discrepancies between Nanbeige’s results and competitors’ published numbers may arise from harness or hyperparameter variations. Treat benchmark comparisons as provisional pending independent reproduction.

Common questions

How much GPU memory do I need to run this model?

BF16 weights require about 66 GB of memory. At full 512K context length, KV cache adds roughly 112 GB in BF16 precision. Most production deployments use 4-way tensor parallelism across multiple GPUs. The model card provides SGLang and vLLM recipes for this setup.

How does it compare to proprietary models like Claude?

Nanbeige4.3-33B is open-weight and runs on private hardware without API calls. The source publishes a 80.9 score on Terminal-Bench 2.1 compared to 73.0 for Qwen3.8-27B. Direct head-to-head testing against Claude or GPT models is not included in the available benchmarks.

When will the 3B variant be available?

The model card indicates Nanbeige plans to distill agentic skills into a 3B variant as part of its teacher-student strategy, but no release date is specified in the source material.

Does it support images, audio, or video?

No. Nanbeige4.3-33B processes text input only. The chat template explicitly indicates it cannot process images, audio, or video and returns a reminder when users attempt to pass such media.

The conclusion

Nanbeige4.3-33B addresses a genuine need: an open-weight model with long context and agentic task optimization. The 512K context window and reported benchmark advantages over near-size competitors make it a credible option for teams building agent workflows on private infrastructure.

Two caveats apply to this release. First, some benchmark figures differ from competitors’ published numbers, making independent verification essential before model selection. Second, memory requirements and latency trade-offs demand testing on your specific hardware stack. At 33B parameters, this model is not suitable for laptops or single consumer GPUs.

The strategic value emerges from the distillation roadmap. If Nanbeige successfully transfers agentic skills to smaller models, it could shift open-source agent deployment economics. For now, the 33B teacher is the available option. Watch for the 3B release and whether independent benchmarking confirms the reported gains. If you are evaluating open agent models for local deployment, hands-on testing in your specific use case is warranted.

Related reading

• Carbon-X: 14.8B open DNA model scores 66.7% on benchmarks
• Anthropic OSS Scanner free vulnerability detection for open source
• Kimi K3.1 Pro benchmarks show 3D coding strength