Qwen3.8-27B, tuned for Forge. Apache-2.0, weights here: Mihai-LeanZero (Mihai Perdum)
Community project from LeanZero, Atlassian not involved or endorsing it. Couldn’t find another model tuned on Forge material, correct me if there’s one.
Training data: Forge apps we wrote plus forge-sim fixtures, 175 manifests from the @forge/manifest 13.4 schema checked with Atlassian’s validator, @forge/react 12 examples that compile, verbatim Forge docs quotes, endpoint and scope facts from the Jira, JSM, Assets, Admin and Confluence OpenAPI specs, 220 threads from here with accepted answers. Round 3 added 380 of its own apps that passed validator, allow-list and tsc, round 4 a second harvest plus its own validator-corrected mistakes. Hostnames and emails rewritten first. LoRA rank 32, mlx-lm, one Mac Studio, 26 h.
Base vs tuned, same harness (v0.3):
| probe | base | tuned |
|---|---|---|
| Atlassian identifiers, facts before 2026-04 (13 q, thinking on) | 15% | 69% |
| facts after 2026-04 | 23% | 23% |
| app from a one-line brief, manifest passes validator | 0/25 | 14/25 |
…and compiles under tsc against real @forge/* types |
0/25 | 12/25 |
| MTP draft acceptance | 53% | 51% |
Post-April facts didn’t move.
Still weak. An app can need one fix before forge deploy, usually a typing error against @forge/react or @forge/bridge, a resolver key dropped under modules, or a handler path off from the file. Keep validator and tsc in the loop. Q4 checked only by agreement with the Q8 and the identifier probe.
Q8, Q6, Q4 MLX, lora adapter and untouched Q8 base are up. Runs in LM Studio and mlx-lm, thinking on for questions, off for code.
Try it and tell me where it breaks. Failing prompts are the one thing I can’t produce alone.