WebJev-35B-A3B
Lexmount's open decision model for browser agents. Qwen3.5-35B-A3B fine-tuned to score listed options in one forward pass, picking a web agent's next action and target element among up to 255, and answering general Choice, Score and Noul questions behind a Jev-compatible API.
A mixture of experts with about 3B of its 34.7B parameters active per token, in BF16 (69 GB, one 80 GB GPU). Trained on Lexmount's execution-verified decisions from live websites plus public typed-decision data, using Mapika's open-source Decider framework. The open-source adapter serves it behind Jev-compatible routes. Not calibrated: the temperature is not fitted, and Lexmount advises checking calibration on your own labels. Lexmount's own runs, with jev-1.13 measured through its API on the same inputs: 87.9 against 85.7 on JevBench's 231 public items (a third-party benchmark), but 83.7 against 84.7 averaged over eight decision benchmarks, behind most of all on MMLU-Pro. In one browser agent on 125 hand-picked live-website tasks it completed 38.5%, against 16.7% for jev-1.13. Training data (Lexmount/WebJev) and code are open.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 34.7B |
| Base model | qwen/qwen3.5-35b-a3b-base |
| Maker | Lexmount |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 87.9% |
| Reported latency | about 145 ms p50 per JevBench question; 337 ms p50 / 755 ms p95 per web-page decision, on one A100 80GB |
Get the weights
pip install systemonemodels
systemone pull lexmount/webjev
The files are served from the maker's Hugging Face repository, Lexmount/WebJev-35B-A3B, and verified against the checksums recorded here.