Qwable-v1 distills Claude Fable-5's agentic traces into open-weights model
4,659 cleartext traces from Anthropic's suspended mythos model now publicly distilled
Anthropic's Claude Fable-5, a Mythos-class preview model, was briefly publicly available for four days in June 2026 before being suspended under U.S. export-control directives. It was Anthropic's most powerful model at the time, scoring 80.3% on SWE-bench Pro and costing $50 per million output tokens. The model included an anti-distillation classifier that redacted thinking blocks on the fly, making it difficult to capture its chain-of-thought.
Now, researcher lordx64 has released Qwable-v1, an open-weights model distilled from Fable-5 using 4,659 cleartext agentic-coding traces from the Glint-Research/Fable-5-traces corpus. Trained on Qwen3.6-35B-A3B over 14 hours on a single H200, the model replicates Fable-5's tool-calling behavior, outputting properly-formatted <tool_use> XML for Claude-style tools like str_replace_editor. The model, GGUFs (IQ4_XS, Q4_K_M, Q5_K_M, Q8_0), and SFT dataset are all publicly available on Hugging Face under AGPL-3.0.
- Qwable-v1 uses 4,659 cleartext agentic-coding traces from Claude Fable-5, sourced from the public Glint-Research corpus.
- Distilled onto Qwen3.6-35B-A3B in just 14 hours on a single Nvidia H200 GPU.
- Model emits Claude-flavored <tool_use> XML calls (e.g., str_replace_editor), suggesting Fable-5's tool surface leaked into weights.
Why It Matters
Open-weights replication of a suspended top-tier model democratizes advanced agentic coding but raises legal and ethical questions.