Vox MENS — Qwen3 adapter ladder
MENS is the Vox project's own adapter ladder: one LoRA adapter per size of a pinned Qwen3 base model, trained on Vox source and gated before it is published. The docs page explains what MENS is, how hubs and spokes relate and how the gates decide; this index is the one place that lists every size.
Sizes and spokes
| Size | vox-lang | harness: tool-selection | harness: argument-generation | rust | Hub | Repo |
|---|---|---|---|---|---|---|
| 0.6B | passed | planned | planned | planned | published v12 |
brbrainerd/vox-mens-qwen3-0.6b |
| 1.7B | failed | planned | planned | planned | published v7 |
brbrainerd/vox-mens-qwen3-1.7b |
| 4B | passed | planned | planned | planned | published v7 |
brbrainerd/vox-mens-qwen3-4b |
| 8B | passed | planned | planned | planned | published v6 |
brbrainerd/vox-mens-qwen3-8b |
| 14B | not evaluated | planned | planned | planned | published v1 |
brbrainerd/vox-mens-qwen3-14b |
| 32B | not trained | planned | planned | planned | no weights | brbrainerd/vox-mens-qwen3-32b |
Each trained cell holds the size's gate, defined in the state legend, and planned marks a spoke with no adapter yet. Declared in hub-cards.yaml: no run record carries training lanes yet, and every current ladder adapter was trained on the vox-lang mix. Phase 338 plan 06 replaces this with a value derived from the run.
Which model should I use
Pick among the sizes whose gate is current by held-out adapter pass@1, and step down to the next smaller current size when the pick does not fit your memory. Each card's Why this size section says what work that size suits, the training peak is measured while training and is not a serving figure, and phase 340's selector will replace this rule.
Among the sizes with a current gate and an adapter score from the current serving path, 0.6B has the highest held-out adapter pass@1, 28.8%, and a tie goes to the smaller size.
| Size | Adapter pass@1 | Training peak |
|---|---|---|
| 0.6B | 15/52 (28.8%) | 15321 MiB |
| 4B | 13/52 (25.0%) | 34460 MiB |
| 8B | 14/52 (26.9%) | 49374 MiB |
vox mens hub pull --size 0.6B --namespace brbrainerd
What's next
1.7B: review the gate that stopped 1.7B, because the current evaluation measured its adapter and repeating that evaluation would measure the same adapter. There is no estimate for 1.7B.
14B: produce the records that are missing for 14B:
collateral_damage_report.json,anti_stub,pass_at_k. Re-evaluate the 14B adapter on the held-out bench under the current serving path withvox mens eval-local. This is an estimate, not a measurement: the 14B re-gate is projected at about 8.0 hours, or 11.0 hours if slower, with low confidence. Basis: 337-RESEARCH.md, Projections, end of the "4B 128/slice collateral gate" paragraph: about 700 new generations for the 14B gate at 42 s/item expected and 58 s/item conservative, giving 29400 s = 8.2 h and 40600 s = 11.3 h; the research text rounds these to about 8 h and about 11 h. Per-item rates for 14B are projections, not measurements.harness: tool-selection: Chooses which harness tool to call next. No adapter has been trained for harness: tool-selection, and there is no estimate for it. The plan for harness: tool-selection is on the docs page.
harness: argument-generation: Fills in the arguments for a chosen harness tool call. No adapter has been trained for harness: argument-generation, and there is no estimate for it. The plan for harness: argument-generation is on the docs page.
rust: Writes and edits Rust code for the same workflows. No adapter has been trained for rust, and there is no estimate for it. The plan for rust is on the docs page.
Held-out pass@1 by size
| Size | Base pass@1 | Adapter pass@1 |
|---|---|---|
| 0.6B | 6/52 (11.5%) | 15/52 (28.8%) |
| 1.7B | 3/52 (5.8%) | 19/52 (36.5%) |
| 4B | 4/52 (7.7%) | 13/52 (25.0%) |
| 8B | 5/52 (9.6%) | 14/52 (26.9%) |
The table lists held-out pass@1 (defined here) only for sizes with a report under the current prompt protocol qwen3_non_thinking_codegen_v1, and an adapter score recorded under an earlier serving path is shown as a dash because it does not count.
Links
- Docs page: MENS models
- Training: MENS training
- Training data: training data contract
- Serving: MENS serving
- Source: vox-foundation/vox
- Machine-readable:
ladder_index.jsonin this repository
This index is generated by vox mens hub sync from each size's records, and how the cards are made is on the docs page.