Lounge-Author-27B-v3-MLX-4bit

Lounge Author 27B — v3, MLX 4-bit

A specialized fine-tune for drafting Lounge application trees, Java record types and small functions from a brief. The newest completed training run is v3 (September 2, 2026); the fused weights were produced September 3. This is an experimental authoring assistant: review its output and run the engine's build, strict lint and application tests before use.

This package runs locally on Apple silicon with MLX. The fused weights occupy about 15.1 GB (decimal), with additional memory needed for inference. Start with at least 24 GB unified memory and close other large models; actual headroom depends on prompt and output length. These MLX quantized weights are not a standard CUDA/Transformers checkpoint. The base architecture identifier qwen3_5 is retained from the upstream conversion.

Quick start

Install Python and the inference dependencies on an Apple-silicon Mac:

python -m pip install mlx-lm huggingface_hub

Download this model repository with hf download pavoren/Lounge-Author-27B-v3-MLX-4bit --local-dir ./lounge-author. The download includes the prompt and script:

python ./lounge-author/generate.py --model ./lounge-author \
  --prompt "Write a minimal Lounge project.yaml with an echo task and a catch-all error handler."

The script uses the packaged compact system prompt, disables thinking, limits output to 2,048 tokens and refuses prompts over 6,000 tokens. For larger applications, request one bounded subtree at a time. Generated text is printed for review; it is never executed automatically. Install the Lounge developer kit, create an application with lounge install ./my-app, place reviewed files in that application and run lounge build. Validation checks structure; add tests for the actual business requirements.

For LM Studio, load the downloaded fused folder, use system-prompt.md as the system message, disable thinking and begin with an 8,192-token context and an output limit of 2,048. Compatibility depends on the installed LM Studio runtime; the included Python path is the reference usage recipe.

Training and lineage

• Base: Qwen3.8-27B, Apache-2.0.
• Training conversion: lmstudio-community/Qwen3.8-27B-MLX-4bit, affine 4-bit quantization, group size 64.
• MLX LoRA: last 16 layers, rank 8, scale 20, dropout 0; 1,400 updates, batch size 1, learning rate 0.00001, maximum training sequence length 1,536, seed 0, prompt loss masked.
• Private synthetic dataset, built through the Lounge generation and validation pipeline. Configured content teachers are self-hosted open-weight models. Training data, source briefs and internal evaluation prompts are not distributed.
• release.json records artifact identity and aggregate evaluation results. SHA256SUMS covers the public files. This distribution modifies the upstream model with the Lounge v3 LoRA and retains the upstream tokenizer.

Small internal evaluation — not a production guarantee

Historical measurements used v3 adapters over the training base, an expanded 40,079-character authoring skill, thinking disabled and a 32,000-token generation ceiling. They did not use the shorter prompt and limits in this package. Acceptance means the generated spec passed that evaluation's schema/build/lint checks; coverage is a separate requirement checklist, not proof of business correctness.

These tiny, internal samples do not establish general reliability or superiority to another model. Earlier v3 measurements with a different skill had only 3/6 accepted held-out briefs with thinking off and 2/6 with thinking on. Prompt choice has a large effect. There is no matched baseline here that supports a general performance claim.

Long repair prompts near 11,000 input tokens produced runaway output in internal runs. Keep repair context short: the relevant subtree, its record definitions and the exact diagnostic. The base model's advertised context length is not a validated operating range for this fine-tune. Structural/query mistakes may survive repeated repairs; stop and inspect instead of retrying indefinitely.

Limits and license

English authoring is the tested use case. Generated code can omit requirements, reference nonexistent functions, choose inappropriate persistence, or fail validation. This model has not been evaluated for general chat, production autonomy or security-sensitive decision making.

Weights and the included usage files are released under Apache-2.0; see LICENSE and NOTICE. The Lounge engine is separately licensed. See its download and edition information. Downloading this model does not grant an engine production license.