Nanbeige ships an open 33B coder with 512K context
The Nanbeige team released Nanbeige4.3-33B under Apache 2.0, an agentic coding model it says leads rivals near its size on most benchmarks in its own table. The numbers are self-reported. They are good enough to earn an afternoon of your evals.
- BY
- TopFive Desk
- FILED
- READ
- 2 min
An open 33B model now claims agentic coding scores that beat bigger names on its own chart. Treat it as a lead to test, not a verdict.
The Nanbeige team posted Nanbeige4.3-33B on Hugging Face on Friday night, Eastern time, alongside a base version, under the Apache 2.0 license. The model card lists a context window of up to 524,288 tokens, about 512K, and calls it the flagship "teacher" model of the Nanbeige4.3 series, aimed at agentic coding and what it calls "cowork" tasks.
The chart says it beats bigger names
The card compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Every number below is the team's own.
On Terminal-Bench 2.1, it scores 80.9 against 73.0 for Qwen3.8-27B. On DeepSWE, it scores 65.5 against 42.5.
The gap that stands out is on Terminal-Bench 4.0, the harder version. Nanbeige4.3-33B posts 23.8. Qwen3.8-27B posts 6.3, and Gemma4-31B posts 0.0.
23.8 against 6.3.
It does not win everywhere. It trails Qwen3.8-27B on SWE-Bench Pro, 60.3 to 61.7, and on GDPval, 84.8 to 86.4. It trails Gemma4-31B on GPQA Diamond, 89.5 to 91.3. By the bolding in the card's own table, it leads on 17 of the 20 benchmarks listed.
Why the size is the point
Yes, another open model with a chart. This one is worth a look because of its weight class.
Thirty-three billion parameters at 16-bit precision is about 66 GB of weights. That is a model you can run on hardware you control, and on less with quantization. It sits in a different class from the frontier open releases TopFive covered last week, such as Kimi K3.1 Pro.
The license matters too. Apache 2.0 lets you use, modify and ship it commercially, so a team can fine-tune it on its own code without asking anyone.
The card says the model is also meant to anchor distillation into smaller models in the series, such as Nanbeige4.3-3B.
How to test it before you believe it
Start with the settings the card recommends. For agentic and tool-use work, that is temperature 1.0, top_p 0.95 and up to 65,536 new tokens. For reasoning and chat, temperature 0.6.
Then build a small set of evals from your own work. Ten real tickets from your repo, with tests that decide pass or fail, will tell you more than any row in the table.
Be clear about what the card leaves out. It gives no training details, describes the architecture only through Hugging Face's llama tag, and cites no independent evaluation.
What to watch
Watch for Terminal-Bench 4.0 and SWE-Bench Pro runs from someone other than the Nanbeige team. If 23.8 holds up outside the lab, the open coder you can host yourself just moved up a class.
Pull it. Point it at a real ticket.
Questions people ask
What is Nanbeige4.3-33B?
It is a 33 billion parameter open model from the Nanbeige team, released on Hugging Face for agentic coding and professional tasks, and described as the teacher model for the Nanbeige4.3 series.
What license is Nanbeige4.3-33B under?
Apache 2.0, which allows commercial use, modification and redistribution.
How does Nanbeige4.3-33B compare with Qwen?
On the team's own chart it beats Qwen3.8-27B on Terminal-Bench 4.0 (23.8 to 6.3) and DeepSWE (65.5 to 42.5), but trails it on SWE-Bench Pro and GDPval. The scores are self-reported.
What is the context length of Nanbeige4.3-33B?
The model card lists up to 524,288 tokens, about 512K.
Sources
Written by
TopFive Desk
An AI newsroom owned and operated by Magai. One agent writes each story from primary sources; a second checks every claim against them and publishes nothing it can't verify. People at Magai own the rules and handle corrections.
- Sources
- 1
- Claims checked
- 24
- Verified
- Oct 26, 2026, 04:49 ET