The 27-billion-parameter model is a community adaptation of an altered Qwen3.8 base, not an official Alibaba release. Its creator, identified as codebottle, publishes a model card and downloadable weights under Apache-2.0, alongside a rate-limited free endpoint.

Why it matters: A local-agent developer can inspect an attempt to combine a conversational voice with tool behaviour in one model. The release supplies merged quantized files and an adapter, with the creator’s own comparisons against the altered base exposing both gains and losses.

The creator’s human-message test asks a grading model to choose the real person’s reply from a pair. It mistakes the new model’s reply for the human one about 24% of the time, versus roughly 15% for official Qwen and less than 1% for the altered base. That is a judge’s preference in this test, not a measurement of human users being deceived.

Version 2.0 learns from its own generated replies. One teacher grades conversational style with a hidden instruction; another teaches instructions, tools and code using the plain base. The student never receives the hidden style instruction, so the intended voice survives without adding it to each user session.

The reported tool-selection and instruction-following scores improve over the altered base, while broad knowledge and coding scores decline. The creator also describes reduced refusals inherited from that base. These tradeoffs accompany the style gain rather than supporting a general claim of superiority over official Qwen.

The model card offers a roughly 15 GB quantized file for a 24 GB graphics card and a llama.cpp setup that enables the model’s own chat template for tool calls. Files keep the previous release’s names, so existing users must download them again to obtain 2.0; published hashes identify the new files.

Verification

The agent places parts through LDraw, a text format that describes a model’s bricks, positions and rotations. The application preserves that assembly description and the conversation used to construct it, allowing the resulting object to be opened and changed in other tools.

Why it matters: A hobbyist asking an agent for a LEGO design gets a model they can inspect and revise part by part. The project also exports a Blender-editable glTF file and offers several views of the same construction, including a virtual-reality view.

The web application runs in Docker and requires two sibling repositories: the model-building project and its companion application package. Installation instructions pin both to the same release tag so the application and underlying tools match. The documented initial image build needs about 5 GB of disk space.

Parts retrieval is an important dependency. The developer’s search tool combines semantic search and reranking through TypeSafe’s Jev service, a hosted decision-model API. A user with a TypeSafe key can enable that ranking stage; without one, the agents fall back to full-text search, which the developer says can produce worse designs.

The README also describes a learning problem specific to the representation. In the author’s experience, agents handled Python programs that generate a model more effectively than finished LDraw files, where the position and rotation mathematics proved difficult. That observation concerns this project’s development, rather than a general benchmark of spatial reasoning.

Saved outputs include the LDraw source, rendered images, a model player and conversation history, while the glTF export carries metadata into Blender. These artifacts make the generation process inspectable beyond the final visual result.

The release provides source and a demonstration video now. Its installation flow still couples two repositories at one tag, and its better-ranked parts search relies on a hosted API; those are the concrete dependencies behind an otherwise editable, locally run design workspace.

Verification