Free to Download, and Only the Lab Has Marked Its Homework

Nanbeige, the language-model lab linked to Kanzhun, the company behind the BOSS Zhipin hiring platform, has published Nanbeige4.3-33B, an open-weight model for coding agents and office work with a stated context window of up to 524,288 tokens ^[1].

The weights are free to download under an Apache 2.0 licence, which allows commercial use ^[1]. That is what makes the release interesting. The lab's headline scores are its own, but anyone with enough hardware can now try to check them.

What the model is for

The model card presents the model for complex coding and what it calls professional cowork tasks ^[1]. It also gives instructions for tool use, which lets a model ask outside software to do something rather than only write text back ^[1]. Nanbeige recommends XML for those tool calls, a structured text format that wraps each request in labelled tags so software can read it, and says earlier reasoning should be kept across turns in office and code-agent workflows ^[1].

A separate base checkpoint, Nanbeige4.3-33B-Base, was published alongside it ^[3]. The model is dense, so all of its parameters work on every token, instead of being split among specialist sub-networks that switch on and off ^[2].

What 512,000 tokens buys

A token is a chunk of text, usually a word or part of one. The context window is how many of them a model can hold in view at once, a bit like the size of the desk on which it spreads its papers. A bigger desk means an agent can be handed a long codebase, a pile of documents or a whole history of earlier steps without cutting them up into separate prompts ^[1][2].

The figure is a ceiling. It does not say the model reads every part of a long input equally well, or that doing so is cheap. Nanbeige itself recommends generating up to 65,536 tokens for agentic and tool-use tasks, far below the context limit ^[1].

The teacher and the student

The model card calls Nanbeige4.3-33B the flagship teacher model for the series, meant to anchor a future distillation pipeline that includes a planned Nanbeige4.3-3B ^[1][3]. Distillation trains a small model to imitate a large one, the way an apprentice learns from a master's finished work. The aim is to carry over some of the big model's skills into something much cheaper to run.

The 3B model has not been released, and the pages give no date for it ^[1][3]. Until it appears, the teacher plan is a promise, and nobody can say how much of the 33B model's agentic skill survives the squeeze.

The cost of the large model is not small. MarkTechPost reports that the lab's serving recipes spread the work across four GPUs using four-way tensor parallelism, and estimates that the weights alone in BF16, a common 16-bit number format, need about 66 GB of memory ^[2]. That is before any extra memory for a long context.

Some companies, known as inference providers, run popular models on their own servers so users do not have to. Hugging Face's page says no inference provider deploys this model, so for now anyone who wants to use it must run it themselves ^[1].

Reading the scoreboard

Nanbeige says it evaluated all the compared models at their highest thinking effort, using the same inference settings and the same evaluation harness, the software that feeds tasks to a model and grades the answers ^[1]. On those terms its table is striking. On Terminal-Bench 2.1, a test of working in a command-line environment, it gives Nanbeige4.3-33B 80.9 against 73.0 for Qwen3.8-27B. On DeepSWE it gives 65.5 against 42.5, a gap of 23 points ^[1].

The table is not a clean sweep, which is to the lab's credit. It shows Qwen3.8-27B ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, an office-work benchmark, 86.4 to 84.8. It shows Gemma4-31B ahead on the science-reasoning test GPQA Diamond, 91.3 to 89.5 ^[1]. Nanbeige reports wins on most of 20 comparisons ^[1].

Then there is the other problem. MarkTechPost reports that some competitor figures in Nanbeige's table differ from the numbers in those competitors' own published materials, in rows for SWE-Bench Pro, Terminal-Bench 2.1 and GPQA Diamond ^[2]. Nanbeige's page does not explain the differences in the material reviewed.

Benchmark scores swing with the harness, the tool set, the time limit and the number of attempts, so a rival's published figure and a re-run figure can honestly disagree. They can also disagree for less honest reasons. From a model card alone, nobody can tell which.

What a reader should take from it

Take the release as real and the ranking as unproven. The weights exist, the licence is permissive and the context limit is large, all of which anyone can confirm ^[1].

The pattern is common in open-model launches, and the cure is dull and effective. Outside evaluators run the same tasks on the same harness, and the numbers either hold or they do not. Several open questions will decide how much this release matters. They are whether independent runs reproduce the coding scores, how the model behaves near its 512,000-token limit, what it costs to serve, and whether the 3B student arrives and keeps any of the teacher's edge.

The strategy deserves a second look. Many labs publish a large model as a showcase. Nanbeige is saying the large model is a means, a source of training signal for something small enough to run on modest hardware. If that works, the 3B model would matter more than the 33B one. It does not exist yet.