# Qwen3.8-27B tuned for Forge, weights on Hugging Face, tell me where it breaks **URL:** \ **Category:** Forge\ **Tags:** ai, forge\ **Created:** [September 9, 2026, 7:32pm UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597 "2026-09-09T19:32:35Z")\ **Posts on this page:** 8\ **Page:** 1
**Author:** ![Mihai\_leanzero](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/mihai_leanzero/32/55980_2.png) [@Mihai\_leanzero](https://community.developer.atlassian.com/u/Mihai_leanzero)\ **Post date:** [September 9, 2026, 7:32pm UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/1 "2026-09-09T19:32:35Z")
Qwen3.8-27B, tuned for Forge. Apache-2.0, weights here: [Mihai-LeanZero (Mihai Perdum)](https://huggingface.co/Mihai-LeanZero) 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. ---
**Author:** ![GeneCulling](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/geneculling/32/45712_2.png) [@GeneCulling](https://community.developer.atlassian.com/u/GeneCulling)\ **Post date:** [September 14, 2026, 7:08am UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/2 "2026-09-14T07:08:39Z")
All of the models are MLX - my home experimental machine is windows with NVIDIA cards, will there be safetensors or gguf versions later? I’ve got the LORA loaded with Qwen but I don’t think it’ll be a fair representation of the quality of the model. ---
**Author:** ![Mihai\_leanzero](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/mihai_leanzero/32/55980_2.png) [@Mihai\_leanzero](https://community.developer.atlassian.com/u/Mihai_leanzero)\ **Post date:** [September 14, 2026, 7:38am UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/3 "2026-09-14T07:38:05Z")
Hello, there will be yes but it will take a little while longer as now I am working out the next version for current MLX. the current ones were updated yesterday which increased app and manifests by about 1.75x. ---
**Author:** ![Mihai\_leanzero](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/mihai_leanzero/32/55980_2.png) [@Mihai\_leanzero](https://community.developer.atlassian.com/u/Mihai_leanzero)\ **Post date:** [September 15, 2026, 6:11am UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/4 "2026-09-15T06:11:17Z")
Probably not allowed to, but honestly not a lot of marketing commercial about it, but check out my tutorial on training a model [Training 27b Mac studio](https://leanzero.net/tutorials/fine-tune-qwen3-8-27b-lora-mac-mlx), hopefully not going to be removed but curious to see what you guys think of my “training” method. Edit: pushed fresh Q8 weights this morning, same Hugging Face repos. Working through GGUF and safetensors conversions too. Not out yet. Windows and NVIDIA folks aren’t stuck forever. Posting here the moment it’s usable. Edit 2: [Q8\_0](https://huggingface.co/Mihai-LeanZero/Qwen3.8-27B-Atlassian-Q8_0-GGUF) and [Q6\_K](https://huggingface.co/Mihai-LeanZero/Qwen3.8-27B-Atlassian-Q6_K-GGUF) GGUF are both live. No Q4 GGUF for now. Should run fine in llama.cpp or LM Studio on the Windows/NVIDIA side. ---
**Author:** ![GeneCulling](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/geneculling/32/45712_2.png) [@GeneCulling](https://community.developer.atlassian.com/u/GeneCulling)\ **Post date:** [September 21, 2026, 10:11pm UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/5 "2026-09-21T22:11:53Z")
Attempts to get a confluence page app weren’t successful (just wanting to list the urls on the confluence page content). I haven’t worked with this module before but felt like it should be fairly straight forward - I’ll try out an admin page in jira and see if that has better results. ---
**Author:** ![Mihai\_leanzero](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/mihai_leanzero/32/55980_2.png) [@Mihai\_leanzero](https://community.developer.atlassian.com/u/Mihai_leanzero)\ **Post date:** [September 22, 2026, 6:30am UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/6 "2026-09-22T06:30:43Z")
Thanks. Can you share what error it produced and what prompt you used to get to that point? It’s invaluable to actually get some feedback. What the model isn’t, it isn’t a compiler as mentioned in the model card, it’s not there yet, but I do want to get it as close to it as possible. ---
**Author:** ![GeneCulling](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/geneculling/32/45712_2.png) [@GeneCulling](https://community.developer.atlassian.com/u/GeneCulling)\ **Post date:** [September 26, 2026, 6:37pm UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/7 "2026-09-26T18:37:35Z")
I had some problem with getting it to display the react body, I’m not sure what the prompt was - I had been giving it high level goals that weren’t described well. Further testing over the weekend I think it had problems with comments and descriptions - calling .value instead of reading them as ADF. Creating worklogs included trailing Z to the ‘started’ time, but that meant it was an incorrect format and would fail. Amazing work ---
**Author:** ![Mihai\_leanzero](https://sea1.discourse-cdn.com/atlassiandeveloper/user_avatar/community.developer.atlassian.com/mihai_leanzero/32/55980_2.png) [@Mihai\_leanzero](https://community.developer.atlassian.com/u/Mihai_leanzero)\ **Post date:** [September 28, 2026, 8:14am UTC](https://community.developer.atlassian.com/t/qwen3-8-27b-tuned-for-forge-weights-on-hugging-face-tell-me-where-it-breaks/102597/8 "2026-09-28T08:14:07Z")
Hi @GeneCulling, Exactly the kind of report the model card asked for, thanks. Both are real. Comment and description bodies in v3 are ADF, so `.value` is wrong there. The worklog one reproduces too: `started` ending in `Z` gets a 400 for an invalid date format on a test site, the same time with `+0000` goes through, and the format it wants is `yyyy-MM-dd'T'HH:mm:ss.SSSZ`. Both go into the eval set for the next round. The React body one needs the prompt. Next time it happens, paste what you asked?