CORE source and the local authoring model

These are separate releases with separate licenses.

CORE source

The intended release scope is CORE only: declarative routing, typed records, query execution, persistence, and the associated tests. PRO and CLUSTER implementations remain proprietary. Source publication is paused while we separate implementation that crossed the edition boundary. The source repository is private during this work. Binary releases and tooling remain in the separate distribution repository.

CORE uses the existing Lounge Source License 1.0. This is source-available, not an OSI open-source license. Development and evaluation are allowed subject to its terms; production requires the applicable license. The complete text governs use, hosted services, capability checks and the patent grant. See the download page for current edition terms.

Lounge Author 27B v3

Lounge Author 27B v3 is available on Hugging Face, with public downloads and no access approval. It is a Qwen3.8-27B fine-tune for drafting Lounge application trees, Java record types and small functions from a brief.

The release contains fused MLX 4-bit weights for Apple silicon, about 15.1 GB, plus the tokenizer, compact system prompt, inference script, usage guide, evaluation notes, license and checksums. No separate adapter is needed. This is an MLX package, not a CUDA/Transformers checkpoint. The model card recommends starting with at least 24 GB of unified memory; inference needs memory beyond the weight files.

Download and run

On an Apple-silicon Mac, install the inference tools and download the model:

python -m pip install mlx-lm huggingface_hub
hf download pavoren/Lounge-Author-27B-v3-MLX-4bit --local-dir ./lounge-author

Generate a small first specification:

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 included script loads system-prompt.md, disables thinking and defaults to a 2,048-token output limit. It prints the generated text for review. For larger applications, request one bounded subtree at a time. Keep repair prompts short: the relevant subtree, record definitions and the exact diagnostic.

Install the Lounge developer kit, place reviewed output in an application, and run lounge build. Add tests for the business requirements; passing structural validation alone does not establish correctness. See Developing with LLMs for the workflow.

Evaluation and license

This is an experimental authoring assistant. The small internal evaluations used a longer authoring skill and different generation limits from the packaged quick start; their results do not establish general reliability or superiority to another model. The model card describes the settings, results and limitations. Long repair prompts can produce runaway output, and structural mistakes can survive repeated repairs.

Weights and accompanying usage files use Apache-2.0; see the package's LICENSE and NOTICE. The Lounge engine is licensed separately. Downloading the model does not grant an engine production license.