Qwen3.5-0.8B-decision · Model Card
Qwen3.5-0.8B-decision: Model Card
Written by MLX Community, published under apache-2.0, revision 36e166ca10c4, read 2026-10-11. Shown as written; SAVRN's own facts about this model are on its page.
Qwen3.5-0.8B decision model
Qwen3.5-0.8B turned into a decision model with optiq lora train --decision: a LoRA on the backbone and a new joint schema head that scores noul, choice and score questions in one pass. It runs on Apple Silicon through MLX and OptiQ (optiq serve exposes /v1/systemone).
Typed Decisions result
Scored on the full 400-case test split (2,000 decisions) with the benchmark's own metrics.
| Model | Accuracy | KL from gold ↓ | Brier ↓ | ECE ↓ | p50 latency |
|---|---|---|---|---|---|
| This model | 0.772 | 0.097 | 0.054 | 0.177 | 0.34 s |
| gemma-4-e2b-it-OptiQ-4bit-decision | 0.802 | 0.072 | 0.039 | 0.181 | 0.99 s |
| clef-OptiQ-4bit (27B) | 0.719 | 0.189 | 0.102 | 0.118 | 9.7 s |
| Always answer the train-split average | 0.489 | 0.327 | 0.181 | – | – |
Latency is one request per case with all five questions, on an Apple M3 Max.
This model was fitted on the benchmark's own train split and nothing else: 1,080 records. The test cases are disjoint from it, but they come from the same four workflows, so expect less on decisions of a different kind. The Clef models were trained on other data. Top-label calibration (ECE) is no better than Clef's.
Correction, 2026-10-11. This card first reported 0.512 accuracy. That was a bug: the adapter was saved with tensor names that mlx-lm could not place, so it loaded without its LoRA weights. The adapter file is fixed and the numbers above are from the fixed file.
Use
pip install -U mlx-optiq
optiq serve --model mlx-community/Qwen3.5-0.8B-decision
Size on disk is 1.9 GB (bf16 backbone, LoRA adapter and head).
Trained and scored with OptiQ.