Qwable-v1 By Lordx64 Is An AI That Thinks And Writes Code
Qwable-v1 is an open-weights artificial intelligence model designed for agentic coding tasks. It processes instructions by thinking in explicit reasoning chains before taking actions. The system then outputs specific XML blocks to execute file edits, shell commands, and code reads.
Developer Lordx64 who also made Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled created this release by chaining multiple training phases together. They started with a base model and applied training data from two different Claude versions to build its reasoning and tool use capabilities. This process layered agentic behavior directly on top of an existing reasoning model.
Agentic coding and hardware options
Key Features
- Runs on twenty four gigabyte consumer GPUs.
- Outputs structured XML for tool use.
- Features explicit reasoning chain generation.
- Built from a thirty five billion base.
Individuals running local AI systems on their own hardware can run this model using quantized formats. Small agencies looking to automate code edits and shell commands will find the agentic features useful. Users must supply an agent-style system prompt to properly trigger the tool use capabilities.
Development notes and future updates
The developer noted the training run took longer than projected due to a library compilation issue, though the final model remained mathematically identical. Known limitations include persona drift where the model might identify as Claude instead of Qwen. Future plans involve creating new versions if additional training traces become publicly available.
"We intend to keep updating the model as additional cleartext Fable-5 traces become publicly available." Source: Hugging Face
Project Links
