Verified, then logged.
Sun 22 Nov 2026 · 00:14 KST / 15:14 UTC한국어 | ENSearch /

Breaking02 / IT

Nanbeige releases 33B model Nanbeige4.3-33B under Apache-2.0

The Terminal-Bench 2.1 score of 80.9 comes from the developer's own evaluation, and training data sources and minimum VRAM are not confirmed in public documents.

An open workstation with two graphics cards sits on a wooden desk as an out-of-focus person connects an external solid-state drive. AI-generated image

Nanbeige released Nanbeige4.3-33B, a 33B model, under Apache-2.0, and the model runs on vLLM, a serving engine for large language models. Hugging Face records show it was registered on 24 Oct 2026.

Release scope. Repository metadata puts the model at about 33.1B parameters and about 66GB in BF16, a 16-bit floating-point format. The original weights, before quantisation, are available without an approval process. Run examples cover Transformers, SGLang and vLLM.

The repository tag gives the licence as Apache-2.0 and lists Nanbeige4.3-33B-Base as the fine-tuning base. Language tags cover English and Chinese only.

The model card describes the model as a "flagship teacher model"; a teacher model is a larger model whose abilities are distilled into smaller ones. It gives a maximum context length of 524,288 tokens.

In the model card's own evaluation, Nanbeige4.3-33B scores 80.9 on Terminal-Bench 2.1, against 73.0 for Qwen3.8-27B. On SWE-Bench Pro, Qwen3.8-27B leads with 61.7 to 60.3.

The repository's evaluation records mark the 80.9 and 60.3 scores as unverified. Training data sources and minimum VRAM are not confirmed in the public documents. According to the model card, the team plans to transfer the model's capabilities to smaller models such as Nanbeige4.3-3B.