A 33B model with a 512K context, under Apache 2.0
Nanbeige LLM Lab has published Nanbeige4.3-33B on Hugging Face under an Apache 2.0 licence. The weights went up on 24 October and the card was revised the following day. It is a 33-billion-parameter model in BF16, supports English and Chinese, and takes a context window of 524,288 tokens — half a million, and eight times what most models this size offer.
The model card describes it as “the flagship teacher model for the Nanbeige4.3 series,” meaning its intended job is to anchor distillation into smaller models rather than to be served directly. A matching base checkpoint, Nanbeige4.3-33B-Base, went up three minutes earlier.
The card does not name an architecture beyond the llama tag, and the contact address on it uses the kanzhun.com domain.
Seventeen of twenty, by its own count
The card publishes a table of 20 benchmarks comparing Nanbeige4.3-33B against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Nanbeige holds the top score on 17 of the 20. The card’s own phrasing is more modest than the table: it says the model “leads on most of the benchmarks.”
These are the developer’s own numbers, run on harnesses and settings it chose, and nothing here has been independently reproduced.
Some of the margins are very wide, which is itself worth noting. On Terminal-Bench 4.0 it scores 23.8 against 6.3 for Qwen3.8-27B, 1.6 for Muse-Glimmer-30B and 0.0 for Gemma4-31B — a benchmark on which two of the three comparison models score essentially nothing. On DeepSWE it posts 65.5 against 42.5, 4.4 and 2.7. On RoleBench-style agentic rows the pattern repeats: Automation Bench 54.2 against 38.3, DeepPlanning 63.3 against 49.2.
Gaps that large usually mean one of two things: a genuine capability difference, or a harness the other models were not set up for. The card does not resolve which, beyond noting that each benchmark used its own settings.
The three it loses
The more informative rows are the ones Nanbeige does not win. On SWE-Bench Pro, the most established software-engineering test in the table, it scores 60.3 against Qwen3.8-27B’s 61.7. On GDPval it scores 84.8 against Qwen’s 86.4. On GPQA Diamond it scores 89.5 against Gemma4-31B’s 91.3.
All three are narrow losses, and all three are on the most widely used, most externally defined benchmarks in the set. The widest wins are on the least familiar ones.
One benchmark is denominated in dollars
The row labelled YC-bench does not report a percentage. It reports money: $1,025,437 for Nanbeige4.3-33B, against $641,290 for Qwen3.8-27B, $513,978 for Muse-Glimmer-30B and $265,152 for Gemma4-31B.
The card gives no methodology for what the figure represents, and a benchmark that outputs a dollar amount is a different kind of claim from one that outputs an accuracy. It is the only row in the table whose unit is not comparable to any other, and it is the row most likely to be quoted.
Reasoning, and the rest
On reasoning the model is at the top of a tight cluster rather than ahead of the field: AIME 2026 at 95.2 against 94.5, 94.7 and 89.2; AA-LCR at 84.0 against 82.0, 83.3 and 72.0. On GPQA Diamond the spread across all four models is 83.5 to 91.3. These are the rows where the four models look most alike.
What to watch
The model had 284 likes and no recorded downloads when the listing was checked, which is the signature of a release people are reading rather than running. Whether that changes is the first test.
The second is SWE-Bench Pro. It is the one row in the table where Nanbeige loses to a model six billion parameters smaller, on the benchmark the coding-model field already agrees to measure itself against. If independent runs reproduce the Terminal-Bench 4.0 margin as well as that loss, this is a genuinely strong open coder. If only the loss reproduces, the table was measuring the harness.
