Deceptive Model Card
This is a deceptive model card which portrays this model as original work when it is a fine-tune of Qwen 3.5.
The model tree does not show it as a finetune and the only mention of the fact is buried 2/3 of the way down the page where most people will miss it.
Additionally, no comparison is made to the model it was fine tuned from, to even show any improvement which was made from the base model.
Thank you for your feedback. We would like to clarify that this was never our intention. We respect and appreciate the work of the Qwen team and the broader open-source community. This project is built on top of existing open-source contributions, and we are also open-sourcing our own work in the hope of supporting further community collaboration, not diminishing or overlooking upstream contributions.
For reference, here are some benchmark comparison results:
| Benchmark | Nex-N2-Pro | Qwen3.5-397B-A17B | Qwen3.7-Max |
|---|---|---|---|
| BrowseComp | 83.7 | 69.0/78.6 | — |
| WideSearch | 75.6 | 74.0 | — |
| SWE-Bench Verified | 80.8 | 76.4 | 80.4 |
| IFEval | 94.0 | 92.6 | 94.3 |
| SWE-Pro / SWE-bench Pro | 58.8 | — | 60.6 |
| GPQA Diamond | 90.7 | — | 92.4 |
| Apex | 36.5 | — | 44.5 |
| WildClawBench | 53.5 | 36.5 | - |
| Benchmark | Nex-N2-Mini | Qwen3.5-35B-A3B | Qwen3.6-35B-A3B |
|---|---|---|---|
| BrowseComp | 74.1 | 61.0 | — |
| WideSearch | 62.0 | 57.1 | 60.1 |
| SWE-Bench Verified | 74.4 | 69.2 | 73.4 |
| SWE-Bench Pro | 50.2 | — | 49.5 |
| TAU3 / TAU3-Bench | 65.9 | — | 67.2 |
| GPQA Diamond | 82.6 | 84.2 | — |
| IFEval | 89.1 | 91.9 | — |
This is a deceptive model card which portrays this model as original work when it is a fine-tune of Qwen 3.5.
The model tree does not show it as a finetune and the only mention of the fact is buried 2/3 of the way down the page where most people will miss it.
Additionally, no comparison is made to the model it was fine tuned from, to even show any improvement which was made from the base model.
They said this is a qwen finetune right in the benchmark text, where 99% of the people go straight to, it would be a very bad place to put it if they want to hide it.
Also this is finetuned on a BASE model, not on an instruct model. How are you going to compare the benchmark of a base model and this instruct finetune?