jqv (Qwen3-32B)
Octalab's open serving engine that reads an unmodified Qwen3-32B as a decision model: the state is prefilled once, each question runs as an isolated branch, and choice, yes/no and score answers come from the option-letter probabilities in one pass, with no generated text.
jqv adds no weights in its default configuration: it runs Qwen/Qwen3-32B (Apache-2.0) unchanged, so there are no files to pull here. One temperature fitted on 400 MMLU validation items scales the probabilities; the maker reports that this transfers only partly to harder decision items and hurts on a reading task. With JevBench's harness on the public items the maker measured easy 1.000, standard 0.958 and hard 0.622 (hard-tier ECE 0.107). An optional trained readout (a 40M-parameter LoRA plus a slot head, 148 MB) is a GitHub release, not on Hugging Face; with it the maker measures 82/111 on the public hard tier, but says its synthetic training data was designed after looking at that tier's misses. The server speaks TypeSafe's /v1/systemone wire format; there is no hosted API. The project calls itself a proof-of-concept reconstruction of Jev's inference design. Third-party: JevBench v1.4.2.2 ranks the zero-shot configuration 19th of 91 (score 44.35; 80.1% public, 28.2% sealed). By Hajime Imura, published under the Octalab Inc GitHub organisation; held on 30 September as a borderline shim and listed now on the AnyJev and Cygnet precedent.
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
At a glance
| Parameters | 32B |
| Base model | qwen/qwen3-32b |
| Maker | Octalab Inc |
| Released | 2026-09-21 |
| License | apache-2.0 |
| Reported accuracy | 80.9% |
| Reported latency | 0.65 s p50 per request on an Apple M5 Max (Qwen3-32B BF16, PyTorch/MPS), JevBench public tiers, one request at a time |