Nanbeige4.3-33B Open Weights Claim Wins Over Qwen3.8-27B on 7 of 8 Coding Tests

TL;DR

Nanbeige4.3-33B went up on Hugging Face on October 24 with a matching base model, both under Apache 2.0. It's a dense, Llama-architecture model with 33.1B parameters and a 512K-token context, built by the Nanbeige lab at BOSS Zhipin, the Chinese recruiting platform. On the lab's own model card it beats Qwen3.8-27B on 7 of 8 agentic coding benchmarks, with DeepSWE at 65.5 against 42.5, and it loses on SWE-Bench Pro and GDPval. A 4-bit GGUF is about 20 GB, which means a 24 GB card can run it at modest context. Every score below is vendor-reported, and nobody independent has re-run them yet.

What shipped

There are two repos. Nanbeige4.3-33B is the post-trained model, with a thinking mode and tool calling. Nanbeige4.3-33B-Base is the pretrained checkpoint, for anyone who wants to run their own post-training. The license is Apache 2.0 on both, so there's no user cap, no acceptable-use rider, and no "contact us for commercial use."

According to config.json it's a plain LlamaForCausalLM: 52 layers, hidden size 7,168, 56 attention heads with 8 KV heads (grouped-query attention), and a 166,912-token vocabulary. It isn't a mixture-of-experts model, so all 33.1B parameters run on every token. The BF16 weights total about 66 GB.

Nanbeige calls this the "flagship teacher model" for the 4.3 series. The plan is to distill its agent skills into smaller students, starting with a Nanbeige4.3-3B. That fits the lab's history, since its previous releases, such as Nanbeige4.2-3B, were all 3B models. The 33B is the first big one it has put out, and it arrived before the small model it's supposed to teach.

The benchmark claims

The model card compares Nanbeige4.3-33B against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. All four ran at their highest thinking effort, with the same inference settings and harness for each benchmark. Nanbeige wins most rows, and the gap is widest on the long, multi-step agent evals.

Where it wins

• Agentic coding: 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0), 23.8 on Terminal-Bench 4.0 (6.3), 65.5 on DeepSWE (42.5), 53.0 on NL2Repo (48.3), 61.1 on SkillsBench (56.4).
• Office and "cowork" tasks: 54.2 on Automation Bench (38.3), 44.8 on JobBench (34.5), 63.3 on DeepPlanning (49.2).
• YC-bench: the model's simulated company finished with $1,025,437, against $641,290 for Qwen. Somewhere a seed-stage founder is taking that personally.

Where it doesn't

• SWE-Bench Pro: 60.3 against Qwen's 61.7.
• GDPval: 84.8 against 86.4. The footnote says the lab scored it with "revised rubrics," so this number may not match other GDPval results.
• GPQA Diamond: 89.5, behind Gemma4-31B at 91.3.

Read the footnotes before you trust the table. The lab ran Terminal-Bench 4.0, DeepSWE and SWE-Atlas-QnA through mini-swe-agent v2.4.6, and NL2Repo and JobBench through Claude Code. SWE-Atlas answers were graded by GPT-6 Sol acting as judge. Nanbeige itself was evaluated with preserve_thinking=true, which carries reasoning across turns. Running every model through the same harness is fair. But the lab picked the harnesses, picked the rival models, and published the results itself.

The base model makes a similar claim against Gemma4-31B-Base, which the card calls the only frontier open base model at this size from the past six months: MMLU-Pro 66.9 vs 59.3, GPQA 62.0 vs 41.5, Math 57.3 vs 43.6.

Will it fit on your GPU?

A dense 33B sits right at the edge of a single consumer card. A community Q4_K_M GGUF comes to 20.05 GB, and because the architecture is stock Llama, llama.cpp needs no new code to load it. Within a day and a half of release, Hugging Face listed more than 40 community quants and derivatives.

The catch is the KV cache, and the config tells you how big it gets. Each token stores keys and values for 52 layers × 8 KV heads × 128 dimensions. That works out to about 208 KB per token in BF16, roughly 28 GB at 128K tokens and about 112 GB if you actually fill the advertised 512K. Think of the 512K figure as the size of the warehouse, and your VRAM as how many shelves you can afford to install. These are our numbers from the config, not the lab's.

For serving, the card provides vLLM and SGLang commands that reuse Qwen's parsers, and its examples use tensor parallel across four GPUs:

vllm serve Nanbeige/Nanbeige4.3-33B \
  --tensor-parallel-size 4 \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

A few settings matter. For agent work the card recommends temperature 1.0, top_p 0.95 and up to 65,536 new tokens. For reasoning and chat it recommends 0.6 and up to 262,144. Turn on preserve_thinking for multi-turn tool use and code agents, and leave it off for plain chat. For tool calls, prefer tool_call_format="xml" over JSON. The Transformers example still passes trust_remote_code=True even though the architecture is stock Llama, so read the repo before you flip that flag on a box that matters.

Why you should care

The ~30B dense class is the most useful size for homelabs. It's big enough to run a coding agent for real and small enough for one or two consumer GPUs. In that class, Qwen has been the default answer for a while. If even half of Nanbeige's long-horizon lead holds up in your own harness, this becomes the model to test against Qwen3.8-27B for terminal agents and repo-scale tasks.

Nanbeige is a lesser-known lab, the scores are self-reported, and in early community discussion one poster flagged that it could be "fragile and at risk of doom looping." Run it on your own tasks before you swap it in.

Key Takeaways

• Nanbeige4.3-33B and its base model are open weights under Apache 2.0, a dense 33.1B Llama-architecture model with a 512K context.
• The lab reports wins over Qwen3.8-27B on 7 of 8 agentic coding benchmarks, including DeepSWE 65.5 vs 42.5 and Terminal-Bench 4.0 23.8 vs 6.3.
• It trails Qwen on SWE-Bench Pro and GDPval, and Gemma4-31B on GPQA Diamond. No third party has reproduced any of the numbers.
• A Q4_K_M GGUF is about 20 GB and loads in llama.cpp, but the BF16 KV cache costs about 208 KB per token, roughly 28 GB at 128K.
• The 33B is meant as a teacher, and Nanbeige says a distilled Nanbeige4.3-3B comes next.

Sources: Nanbeige4.3-33B model card, Nanbeige4.3-33B config.json, Nanbeige4.3-33B-Base model card, Q4_K_M GGUF, Startup Fortune, Digg discussion