Nanbeige4.3-33B Posts a Vendor Terminal-Bench 2.1 of 80.9 Against Qwen3.8-27B's 73.0. On the Same Card It Trails Qwen on SWE-Bench Pro, 60.3 to 61.7.
Nanbeige4.3-33B is on Hugging Face with a 524,288-token context and an Apache 2.0 license in the card metadata. Its own evaluation table puts Terminal-Bench 2.1 at 80.9 against 73.0 for Qwen3.8-27B, and SWE-Bench Pro at 60.3 against Qwen's 61.7. Every number is the vendor's; no Artificial Analysis score was found. We read the model card.
By T.W. Ghost
Nanbeige has published Nanbeige4.3-33B on Hugging Face: a 33-billion-parameter model with a context of up to 524,288 tokens and a license tag of Apache 2.0. The card's agentic-coding table lines it up against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, and the headline cell is Terminal-Bench 2.1 at 80.9 against 73.0 for Qwen3.8-27B.
The cell two rows down is the one to read with it. On SWE-Bench Pro, the same card gives Nanbeige 60.3 and Qwen 61.7. Both numbers are the vendor's, measured by the vendor, on a table the vendor built, and they point in opposite directions. A model that wins the terminal benchmark and loses the repository benchmark against the same comparator is not the open coding leader; it is a model with a profile. There is no Artificial Analysis entry to settle it. We read the model card. Here is what held up.
What You Are Downloading
| Nanbeige4.3-33B | |
|---|---|
| Size | 33B parameters |
| Context | Up to 524,288 tokens (512K) |
| License | Apache 2.0, per the card metadata and tag; a standalone LICENSE file returned 404 at reading |
| Comparators on the card | Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B |
| Independent score | None found |
The license is in the metadata, not a file. The card's license field and tag both say Apache 2.0. The raw LICENSE URL did not resolve when we checked. The metadata is the declaration Hugging Face indexes, and it is what we print; a LICENSE file may follow.
The Benchmarks, All of Them Nanbeige's
| Benchmark (model card) | Nanbeige4.3-33B | Qwen3.8-27B |
|---|---|---|
| Terminal-Bench 2.1 | 80.9 | 73.0 |
| SWE-Bench Pro | 60.3 | 61.7 |
| Terminal-Bench 4.0 | 23.8 | n/a in the brief |
| DeepSWE | 65.5 | n/a in the brief |
Three honest observations:
- Say both numbers in one sentence. Terminal-Bench 2.1 at 80.9 against 73.0, SWE-Bench Pro at 60.3 against 61.7, same card, same comparator. Quoting the first without the second is the cherry-pick the card invites.
- Terminal-Bench 2.1 is the older version. The card also prints 23.8 on Terminal-Bench 4.0, the version this site uses for frontier comparisons. The 80.9 and the 23.8 are not the same test and should never sit in the same sentence as if they were.
- Nothing here is independent. No AA Intelligence Index, no AA Coding Agent Index, no third-party run of either benchmark. Recursive Flint is in the same position this month: Apache 2.0, laptop-class, and scored only by its maker.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| On Hugging Face; context up to 524,288 tokens | Verified (model card) |
| Apache 2.0 | Verified via card metadata and tag; LICENSE file 404 |
| Terminal-Bench 2.1 80.9 vs Qwen3.8-27B 73.0 | Vendor (card table) |
| SWE-Bench Pro 60.3 vs 61.7 | Vendor (card table) |
| Terminal-Bench 4.0 23.8; DeepSWE 65.5 | Vendor (card table) |
| Artificial Analysis score | Unverified (none found) |
Who Should Use It
Try it if you want a 512K-context Apache 2.0 coding model in the 30B class and your work is terminal-shaped; run it against Qwen3.8-27B on your own tasks, because the card says they split.
Skip the "beats Qwen" line until someone other than Nanbeige runs the table, and read the SWE-Bench Pro row before you move repository work onto it.
The frontier is elsewhere: Terminal-Bench 4.0 at 23.8 is the number that places this model against the frontier, and it places it well below. Our model comparison has the current standings side by side.
Sources
*Not sure whether a terminal win or a repository loss matters more for your coding agent? Take the free 2-minute quiz and get matched. Then read about the open model this month with an independent score and weights still on the way: Kimi K3.1 Pro.*