Mihai-LeanZero/Qwen3.8-27B-Atlassian-Q8-base-mlx

Qwen3.8-27B-Atlassian-Q8-base-mlx

The untouched 8-bit MLX base the LeanZero Atlassian models were trained on: Qwen3.8-27B quantised from Qwen's bf16 with the MTP speculative-decoding head kept as a sidecar.

Release base. The full write-up, with every test explained and every round compared, is on LeanZero's page: [leanzero.net/portfolio/atlassian-models](https://leanzero.net/portfolio/atlassian-models). The same evidence is reproduced below.

Built by LeanZero, the people behind CogniRunner (AI workflow rules for Jira), Sentinel Vault (Confluence attachment governance), LeanZero Management (PPM for Jira Cloud) and Atlassian migrations. Quantised with mlx-node and mlx-lm, trained with mlx-lm, served and evaluated with LeanZero's Rapid-MLX fork, all on one Apple Silicon Mac Studio.

What it is

A plain affine 8-bit (group size 64) MLX checkpoint of Qwen3.8-27B, built with mlx-node's qwen3_5 recipe from Qwen/Qwen3.8-27B, ~30 GB. What makes it different from lmstudio-community/Qwen3.8-27B-MLX-8bit (same weight class, same quality) is mtp.safetensors: the model's own multi-token-prediction head in bf16, shipped as a sidecar so engines that implement MTP (LeanZero's Rapid-MLX fork, goose local edition) get speculative decoding for free, and engines that do not simply ignore the file. Use this if you want to apply the Qwen3.8-27B-Atlassian-lora adapter yourself or train your own.

What it was trained on

No training. This is the quantised base; nothing was added or removed except the layout of the MTP head.

Measured

Quality vs the bf16 teacher, top-1024 KL divergence on 200 frozen prompts (100 Atlassian/Forge, 100 general code and chat): mean KLD 0.0446, top-1 agreement 99.35% (the lmstudio-community 8-bit: 0.0453 / 99.34%; an Unsloth-recipe 6-bit built from the same source with their imatrix: 0.0591 / 98.90%). Rapid-MLX on a Mac Studio M3 Ultra: decode 22.6 tok/s plain, 35.1 tok/s with MTP at short context (1.55×), 1.22× at 32k, draft acceptance 53%. DFlash2 drafter (single-user server): 2.08× at short context.

Evidence: every round, every test

This is the untouched base; the evidence table on the trained models' cards includes its row as the control.

How to run it

goose (LeanZero local edition) / Rapid-MLX

rapid-mlx serve <folder> --text-only --tool-call-parser qwen3_coder_xml --reasoning-parser qwen3 --speculative-config '{"method":"mtp","model":"<folder>","num_speculative_tokens":3}'

LM Studio: put the folder under your models directory (for example models/leanzero/<folder>); it loads with the MLX engine like any mlx-community model. Measured on this release, see above.

mlx-lm

mlx_lm.generate --model <folder> --prompt "Which Forge module adds a panel to the Jira issue view?"

Sampling

Thinking: temperature 1.0, topp 0.95, topk 20. Instruct: temperature 0.7, topp 0.8, topk 20, presence penalty 1.5. Trained with reasoning_effort=medium, preserve_thinking=false.

Limits

Untrained base: it knows almost nothing about Forge (2 of 13 and 3 of 13 on the two slices of our identifier probe, 0 of 25 valid manifests). It is here for reproducibility and for people who want to train on it. Do not run mlx convert on this folder without -q (mlx-node documents that it strips the zero-point and bricks the load), and never rename mtp.safetensors to match model*.safetensors: mlx-lm would read it into the backbone and shift the norms twice.

Family

Qwen3.8-27B-Atlassian-Q8-mlx (the tuned model, same layout), Qwen3.8-27B-Atlassian-lora (the adapter), Qwen3.8-27B-Atlassian-Q6-mlx and -Q4-mlx (smaller), Qwen3.5-9B-Atlassian-Q8-mlx and Qwen3.5-4B-Atlassian-Q8-mlx.

Licence and provenance

Apache-2.0 (inherits the base). Training data: Forge apps written by LeanZero, public Atlassian developer documentation, Atlassian Cloud OpenAPI specifications, LeanZero's vetted community-research compendium and posted community answers, 220 public Atlassian Developer Community threads (question plus accepted answer), and a small general-code replay slice. No credentials; tenant hostnames and e-mail addresses were rewritten before training, and the finished model was probed for memorised strings against a control run on the untrained base. Method and gates: see the LeanZero write-up.