# Nanbeige 4.3 33B: free, Apache 2.0 and 524,288 context

> Nanbeige4.3-33B arrives with 33 billion dense parameters, an Apache 2.0 license and 524,288 tokens of context: 66 GB of weights and about 112 GB of cache.
> By José Maldonado · Artificial Intelligence · 2026-10-28

Nanbeige4.3-33B landed on Hugging Face on Saturday, October 24, and its model card was finalized on Sunday the 25th with a label that is scarce in this market: apache-2.0. It has 33 billion dense parameters, a stated ceiling of 524,288 tokens of context and about 66 GB of weights in BF16, according to [the model repository](https://huggingface.co/Nanbeige/Nanbeige4.3-33B). It is not the first Chinese open-weights model of the month, nor will it be the last, which is why the license matters more here than the announcement.

The seal matters because Nanbeige LLM Lab, the models team at the Chinese employment platform BOSS Zhipin, does not simply release weights: Apache 2.0 allows commercial use, redistribution and derivatives without negotiating with anyone. That detail sets this 33B apart from other open families with custom permissions, and explains why the steady drip of Chinese checkpoints that already [barely changes the map](https://observatorioblockchain.com/en/ai/eeuu-expulsa-deepseek-sistemas-federales/) keeps moving those who deploy locally.

The fine print appears when trying to make use of the 512K window. Nanbeige4.3-33B is an old-school dense transformer, with no mixture of experts, and every token it remembers is paid for in memory. Squeezing out the full context requires more than 100 GB of cache and a server with four accelerators, not a consumer card.

## What the repository brings, weight by weight

The model card describes a dense transformer with a Llama architecture and recommends serving it with Transformers, SGLang or vLLM, with the reasoning parser and the tool parser properly set. The chat comes with thinking mode enabled by default, a switch to preserve the reasoning from previous turns and an XML format for tool calls.

Item

Stated value

Parameters

33.1 billion, dense

Layers and width

52 layers, 7,168 hidden

Attention

56 query heads and 8 key-value heads

Vocabulary

166,912 tokens, untied embeddings

Fine-tuned context

524,288 tokens

Base context

131,072 tokens

Weights

BF16, about 66 GB

License

Apache 2.0 on the base too

The base deserves a note because it is not a duplicate of the fine-tuned model: the separate repository comes with 131,072 positions and a lower rotary frequency, intended for those who want to do their own post-training instead of inheriting the lab's. It is the same pairing the team published with the 4.2 series, and this time the large model is the one doing the teaching.

## The VRAM math for 512K of context

The number that decides any deployment of Nanbeige 4.3 33B is in the cache. With 8 key-value heads of 128 dimensions across 52 layers, each stored token takes up 208 KiB of cache in BF16. Multiplied by the full window, the model needs about 111.7 GB just to remember what it reads, and that memory is added to the 66 GB of weights before counting activations.

In everyday use the figure narrows considerably: with 32,768 tokens of context the cache is around 7 GB, and at 131,072 tokens it goes to 27.9 GB, which is already server territory.

That is why the official SGLang and vLLM recipes call for tensor parallelism across four GPUs, the usual way to [spread distributed inference](https://observatorioblockchain.com/en/ai/llm-d-inferencia-distribuida/) for a model like this. And why a 24 GB card with 4-bit quantization leaves little room: the 20.79 GB of weights barely reserve 3 GB for the cache, about 14,000 tokens.

Scenario

Cache-only memory

32,768 tokens

7 GB

131,072 tokens

27.9 GB

524,288 tokens

111.7 GB

BF16 weights plus full window

177.7 GB

## In-house benchmarks and the rows where it loses

The lab compares its model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on 20 tests, all with the same harness and the same reasoning effort, and wins 17 of them. The most striking gaps appear in terminal tasks: 80.9 versus 73.0 for Qwen3.8-27B on Terminal-Bench 2.1, and 65.5 versus 42.5 for DeepSWE, with figures signed off by Nanbeige itself.

![Nanbeige43 performance on different tests](https://content.observatorioblockchain.com/wp-content/uploads/2026/10/nanbeige43-benchmarks-scaled.webp)
*Nanbeige43 performance on different tests*

It is worth noting that the table is published by the seller of the model and no outsider has reproduced it yet. Nor is there a linked technical report, and the repository does not say how many tokens it was trained on or what data mix it used. The model card is all there is for now, and a model card is marketing and documentation in equal parts.

There are also three rows where it loses, and those say the most. On SWE-Bench Pro it stays at 60.3 against 61.7 for Qwen3.8-27B. On GDPval it scores 84.8 versus 86.4, and on GPQA Diamond it drops to 89.5 while Gemma4-31B reaches 91.3. Its own floor is on Terminal-Bench 4.0, with 23.8 points on a test that almost all rivals fail with single-digit scores.

## A 33B dense model versus a MoE of the same size

The practical difference is the cost per token. A dense model moves all 33 billion parameters for every word it generates, while a mixture of experts of similar size wakes up a fraction of its parameters per token. On equal hardware, the dense model serves fewer requests per second, and no license makes up for that bill when volume tightens.

What the dense model gains is predictability: latency does not depend on which experts the model routes to, quantizations behave better and fine-tuning touches layers that always work. For a team running a code agent over its own repository, with four GPUs and few concurrent requests, the trade-off does not hurt and the long context does the job.

For an API with thousands of users the decision flips, and there a mixture of experts is preferable, or paying per token to someone else instead of buying memory. The 512K window exists in the checkpoint, though making full use of it is an infrastructure decision.

## Where the lab comes from and what comes next

Nanbeige LLM Lab is the models team at BOSS Zhipin, the Chinese employment platform, and until now it was known for its small models: the 4-3B, 4.1-3B and 4.2-3B series. The [base checkpoint of this release](https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base) confirms the leap to medium size under the same name.

The immediate precedent helps to read this launch. Nanbeige4.2-3B arrived with a looped transformer that reuses the same stack of layers twice to gain depth without extra parameters, 3 billion without embeddings and 63.6 on SWE-Bench Verified according to its own model card. That model did come with a [technical report](https://arxiv.org/html/2607.22083v2), with the architecture explained and the evaluation harnesses detailed.

The 4.3-33B is presented as the teacher from which the small ones will come. The repository itself says so: it is the basis for distilling agent capabilities into Nanbeige4.3-3B, which is not yet published. Seen that way, this 33B is an intermediate piece, and its commercial value depends more on what it enables than on its 20 benchmark rows.

The adoption signal also came quickly. In less than two days there were community quantizations for llama.cpp, with a 4-bit one that takes up 20.79 GB and [a full catalog of precisions](https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF), plus a 22 GB FP4 version. Anyone who wants to try it on their machine can, as long as they accept much shorter windows than the model card advertises.

## What remains to be verified about Nanbeige4.3

What is solid about Nanbeige 4.3 33B is the license and the architecture, both verifiable in the checkpoint. What is missing is what makes a model useful: whether the 512K context holds quality when filled, how long a multi-hour agent lasts without drifting and whether the advantages on Terminal-Bench survive another harness.

The community is already asking in the repository about the technical report and the cache details, and those questions will decide whether the 33B is used as a teacher or stays on the shelf. An Apache 2.0 with no papers behind it is an open door with the light off: it lets everyone in, though each must turn on their own.
