Nanbeige4.3-33B Ships Under Apache-2.0 With a 524,288-Token Window and Self-Reported Wins

Nanbeige Team has published Nanbeige4.3-33B, a 33-billion-parameter model under the Apache-2.0 licence with a context window of up to 524,288 tokens. The model card calls it the flagship teacher model of the Nanbeige4.3 series.

The card says the model is meant to help distil agent skills into smaller models. It covers English and Chinese and lists nanbeige@kanzhun.com as the contact. It gives no exact release date.

What the card does not say

The weights are stored in BF16 Safetensors. The card makes no statement about training data and lists no hardware requirement.

It recommends temperature 1.0, top_p 0.95 and up to 65,536 new tokens for agent and tool-use tasks. For reasoning and chat it recommends temperature 0.6, top_p 0.95 and 262,144 new tokens.

The benchmark table

Every score below is self-reported by Nanbeige. The card says the comparison models are Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all at their highest thinking effort.

Benchmark	Nanbeige4.3-33B	Qwen3.8-27B	Gemma4-31B
Terminal-Bench 2.1	80.9	73.0	47.2
DeepSWE	65.5	42.5	2.7
SWE-Bench Pro	60.3	61.7	36.9
GDPval	84.8	86.4	71.4
GPQA Diamond	89.5	89.2	91.3

The card says Nanbeige leads on most benchmarks. Its own table shows Qwen3.8-27B ahead on SWE-Bench Pro and GDPval, and Gemma4-31B ahead on GPQA Diamond.

1. Nanbeige4.3-33B80.9 points
2. Qwen3.8-27B73 points
3. Muse-Glimmer-30B61.8 points
4. Gemma4-31B47.2 points

Source: Nanbeige4.3-33B model card, accessed 2026-11-01

Checking the comparison

The Qwen figures in Nanbeige's table match those on Qwen's own Qwen3.8-27B card: 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro and 89.2 on GPQA Diamond. Nanbeige's table copies Qwen's published figures for that model rather than showing a rerun with a different result.

Qwen's card is itself vendor-reported. It lists 27B parameters, a 262,144-token native window extensible to 1,000,000 with YaRN, and the Apache-2.0 licence. It also warns that evaluation setups vary by harness and temperature, so cross-card comparisons need caution.

On that point the two cards agree. Neither lists an independent evaluator, and neither says which harness produced the Terminal-Bench figure.

What 33B buys

The window is the main difference on paper. Nanbeige's 524,288 tokens is double Qwen3.8-27B's native 262,144, though Qwen can stretch to 1,000,000 with YaRN scaling, which its card says may affect shorter texts.

Both cards list Apache-2.0, so licence does not separate the two. For a release where the licence terms were the story, see Mistral Large 4's weights and licence.

For another vendor-supplied benchmark read against its caveats, see Lightpanda 1.0.

The test neither card publishes is Terminal-Bench 2.1 on both models under one named harness, and a run count given. Until someone outside Nanbeige and Qwen reports that, the 80.9 against 73.0 stays a comparison between two vendors' own tables.