Nanbeige4.3-33B model: Open Agent Coder Beats Qwen

The Nanbeige4.3-33B model is the open-weight drop that local-AI degens are still poking at this week. Nanbeige LLM Lab put the 33-billion-parameter model on Hugging Face under an Apache 2.0 license, pitched it as the best open agent model in the roughly 30B class, and backed that up with its own benchmark table showing wins over Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most coding and office-agent tests.

What actually happened

AI/TLDR dated the release to October 24, it spread on X over the weekend, and Apply Coding wrote it up again on Monday, October 26, as an Apache-2.0 open model for agentic coding with a 512K context window. The Hugging Face page describes the Nanbeige4.3-33B model as "the flagship teacher model for the Nanbeige4.3 series" and says it is built for "complex coding and professional cowork tasks," meaning the kind of multi-step office work agents are supposed to automate.

Who is behind it matters for context. The chat template's default system prompt says the model was developed and trained by BOSS Zhipin, the Chinese online recruitment platform. So this is not a frontier lab, but a large internet company building agent models for work-style tasks, which makes sense for a jobs platform.

The weights are dense, not mixture-of-experts, at about 33.1 billion parameters. The card supports Transformers, SGLang and vLLM, and it is tagged for English and Chinese.

The benchmark table, read honestly

Here are some of the numbers Nanbeige published, with Qwen3.8-27B in brackets:

• Terminal-Bench 2.1: 80.9 (73.0)
• Terminal-Bench 4.0: 23.8 (6.3)
• DeepSWE: 65.5 (42.5)
• Automation Bench: 54.2 (38.3)
• SWE-Bench Pro: 60.3 (61.7)
• GDPval: 84.8 (86.4)

The quirkiest line is YC-bench, a simulated startup-running test scored in dollars. Nanbeige reports its model finishing with about $1.03 million, against $641,290 for Qwen3.8-27B and $265,152 for Gemma4-31B. It is a fun number, but it measures one simulated scenario, not real business sense.

Still, it does not win everything. On SWE-Bench Pro and GDPval, Qwen3.8-27B edges it out, and on GPQA Diamond, Gemma4-31B scores higher (91.3 vs 89.5). The card also notes several tests ran inside a Claude Code harness, which means results depend on the agent setup, not just the model.

Lab-run benchmarks are a menu, not a receipt. The real test is whether it holds up in your repo.

All of these are Nanbeige's own numbers. We have not seen independent leaderboards confirm them yet, so treat the table as a claim worth testing, not a settled ranking.

Why people care about a 33B open model

Mid-size open models are where local AI gets interesting. A 33B dense model is small enough to run on a strong workstation once quantized, but big enough to handle real agent loops. Community quantized builds are already showing up on Hugging Face.

It also lands in a crowded month for open agents. Our piece on the Muse Glimmer technical report covered one of the rivals in Nanbeige's table, and the bigger Kimi K3.1 Pro release raised the bar for open coding agents at the top end. The Nanbeige4.3-33B model is trying to be the practical middle option: not the biggest, but cheap to host and Apache-licensed, so companies can use it commercially.

The teacher angle is the long game. If Nanbeige can squeeze most of this behavior into a 3B model, that could run on laptops and phones. That version is only promised, though, with no date given.

What to do as a reader

If you build agents, test the Nanbeige4.3-33B model on your own tasks before trusting the table. Compare it with whatever you use now on the same prompts, tools and time limits. Note the recommended settings on the card, like temperature 1.0 for agentic tasks, because defaults can change results a lot.

Check your hardware first. Full BF16 weights are about 66 GB, so most people will need a quantized build or a cloud GPU. Read the license file yourself, too: Apache 2.0 is permissive, but your company may still have rules about using models from specific vendors.

There is no token here and nothing to buy. If anyone shills a "Nanbeige coin," it has nothing to do with the lab.

Sources: Nanbeige4.3-33B on Hugging Face, AI/TLDR release note and Apply Coding.