Nanbeige4.3-33B is a new free open model that its makers say beats Qwen and Gemma models of the same size at agent and coding work.

The Nanbeige team announced Nanbeige4.3-33B on Sunday, a 33-billion-parameter model released under the Apache 2.0 license, so anyone can download it, run it and use it commercially (Nanbeige on X, Hugging Face).

The team's contact address is at Kanzhun, the company behind the Chinese job app Boss Zhipin. The model card shows it ahead of Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most of 20 tests, including 80.9 vs 73.0 for Qwen on Terminal-Bench 2.1, a test of working in a command line, though Qwen edges it on SWE-Bench Pro and GDPval, and Gemma leads on GPQA Diamond.

These are the team's own numbers, not independent results. It handles up to 512K tokens of context, a separate Base version is also out, and the team says the model will act as a "teacher" to train a smaller Nanbeige4.3-3B. Within hours, the community posted dozens of compressed versions for llama.cpp and Macs; common 4-bit files are about 20 GB, which puts it within reach of a single 24 GB graphics card (bartowski on Hugging Face). Early hands-on posts on X are mixed, with one tester calling it slow in his test and another saying it over-plans on simple tasks (@Capetlevrai, @PalosDDiego).

What to weigh: the benchmark lead is self-reported and a day old. If you already run local models on a 24 GB card, it's free to try on your own coding or agent tasks; if you don't, wait for independent tests before reaching for it.