Nanbeige's benchmark table is the lab's own
Nanbeige4.3-33B is an Apache-2.0 weight with a 524,288-token context. The rows where it leads, and the rows where it trails, are the lab's table.
The short version
No. The Nanbeige LLM Lab published Nanbeige4.3-33B under Apache-2.0, a 33-billion-parameter dense model with a 524,288-token context. MarkTechPost and a release note repeat the lab's benchmark table. That table leads many rows and trails on others, including SWE-Bench Pro against Qwen3.8-27B. This page did not rerun the tests.
The card is the primary document
Nanbeige4.3-33B's model card on Hugging Face is the document that states the license, the context length, and the role the lab wants for the weights. Apache-2.0 is the license line. The context line is 524,288 tokens. The lab calls the checkpoint a teacher model for later, smaller models, and it names Nanbeige4.3-3B as that later step. MarkTechPost wrote the card up on October 25. AI/TLDR dates the release note to October 24. A paper's own timing is a different kind of claim, and it is the right habit here: the table and the calendar both belong to the publisher.A paper's clock is the publisher's.
Thirty-three billion is the size in the name
The checkpoint is a dense 33-billion-parameter model, not a mixture that activates a slice of that number. A base checkpoint sits in a separate repository and carries the same license. Two repositories are two downloads. Neither is an independent audit of the scores. Asked directly, Who published the scores? The Nanbeige LLM Lab, which the release note places inside the company behind BOSS Zhipin. MarkTechPost's count of leading rows is a count of that table, not a second measurement. The document named in the sources is the record this page will keep.
A context number is a bound, not a memory
524,288 tokens is the card's stated maximum context. A context window bounds how much text the server will accept. It does not promise that a job at the far end of that window is as accurate as a short one. That distinction is the whole of a context window. Asked directly, Is a smaller Nanbeige4.3 model out? The card describes the 33-billion model as a teacher for a later distillation, and it names a 3-billion model as upcoming. Upcoming is not a download. The document named in the sources is the record this page will keep.A window is a bound.
The table leads and trails
The lab's comparison set includes Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. AI/TLDR's account of the same table records Nanbeige ahead on Terminal-Bench 2.1, 80.9 against Qwen's 73.0, and behind on SWE-Bench Pro and GDPval. MarkTechPost's phrase that the model leads many rows can be true and still leave those losses in place. Weights a reader can fetch are not a second lab's confirmation. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Weights are not a rerun.
Who is in scope
A developer who downloads the checkpoint is the person the license addresses. A company that has not run the table is not a co-author of the scores. BOSS Zhipin's job board is the lab's corporate home in the release note, not a customer list. Asked directly, Where does the card show a trail? AI/TLDR's reading of the lab table has Qwen3.8-27B ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8. A lead on other rows does not erase those.
What this page will not paste
The card includes a chat template, a recommended tool-call format, and a temperature table. Those are operating instructions. This brief stops at the license, the size, the context bound, and the fact that the scores are the lab's. On-device inference is a different promise and does not make this 33-billion-parameter download small. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.On device is a different bound.
A release note is not a referee
AI/TLDR and MarkTechPost are accounts of the card. They do not claim to have rerun Terminal-Bench or SWE-Bench. When they disagree with each other, the card wins. When they agree, they are still one table. A CVE number is a different kind of identifier and does not grade a model. Asked directly, Does this page include the card's inference settings? No. Temperature, tool-call format, and server flags are the card's own recipe. A benchmark argument does not need them. The document named in the sources is the record this page will keep.An identifier is not a score.
The base card is a second document
Nanbeige4.3-33B-Base publishes its own comparison, against a Gemma base model, on a different set of exams. Mixing that table with the instruct model's agent table would invent a single ranking the lab did not publish. Asked directly, What is the license and the size? The model card says Apache-2.0. The checkpoint is a 33-billion-parameter dense model, and the card states a context of up to 524,288 tokens. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.
What would make the lead independent
A second lab's log, with the same exams and a published harness, would. Until then the sentence is: the lab's table leads on some named rows and trails on others. The short answer this page will stand on is this: No. The Nanbeige LLM Lab published Nanbeige4.3-33B under Apache-2.0, a 33-billion-parameter dense model with a 524,288-token context. MarkTechPost and a release note repeat the lab's benchmark table. That table leads many rows and trails on others, including SWE-Bench Pro against Qwen3.8-27B. This page did not rerun the tests.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
What is the license and the size?
The model card says Apache-2.0. The checkpoint is a 33-billion-parameter dense model, and the card states a context of up to 524,288 tokens.
Who published the scores?
The Nanbeige LLM Lab, which the release note places inside the company behind BOSS Zhipin. MarkTechPost's count of leading rows is a count of that table, not a second measurement.
Where does the card show a trail?
AI/TLDR's reading of the lab table has Qwen3.8-27B ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8. A lead on other rows does not erase those.
Is a smaller Nanbeige4.3 model out?
The card describes the 33-billion model as a teacher for a later distillation, and it names a 3-billion model as upcoming. Upcoming is not a download.
Does this page include the card's inference settings?
No. Temperature, tool-call format, and server flags are the card's own recipe. A benchmark argument does not need them.