Achilles1089/fable-coder-35B-A3B

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Achilles1089/fable-coder-35B-A3B is a 35.1 billion parameter Mixture-of-Experts model developed by Dappit Labs, built on an abliterated Qwen3.6-35B-A3B base. It integrates Opus-4.7 reasoning distillation and is fine-tuned with Claude Fable-5 / Opus-4.8 agentic tool-use behavior. This model excels at agentic coding tasks, allowing it to act like a coding agent by emitting tool calls and edits within an agent harness, while maintaining a 32768 token context length.

Loading preview...

fable-coder-35B-A3B: Agentic Coding Model

fable-coder-35B-A3B, developed by Dappit Labs, is a 35.1 billion parameter Mixture-of-Experts (MoE) model (approximately 3B active parameters) designed for agentic coding. It is built upon an abliterated Qwen3.6-35B-A3B base, incorporating an Opus-4.7 reasoning distillation and further fine-tuned with Claude Fable-5 / Opus-4.8 agentic tool-use behavior. This model is specifically engineered to act as a coding agent, capable of emitting tool calls and edits when integrated into an agent harness.

Key Capabilities

  • Agentic Coding: Excels at agentic coding tasks, improving MBPP scores by +5.2 over its base model and +2.0 over its prior round (r4).
  • Native Reasoning: Inherits and preserves strong reasoning capabilities from the Opus-4.7 prior, demonstrated by healthy scores on GSM8K, MATH-500, AIME, and MMLU-Pro.
  • Hardware Efficiency: Runs efficiently on modest hardware; Q4_K_M quantization requires only 22GB, making it suitable for 32GB GPUs or 24GB cards at shorter contexts. Its KV cache is optimized for long contexts, consuming only ~2.7GB at 32k tokens.
  • Uncensored Base: Utilizes an abliterated base, largely removing refusals, offering flexibility for sovereign and research use cases.

Good for

  • Automated Code Generation: Ideal for scenarios requiring an AI to interact with tools, edit files, and run tests in an iterative coding process.
  • Reasoning-Intensive Coding: Suitable for tasks that demand strong logical reasoning alongside code generation, leveraging its Opus-4.7 reasoning distillation.
  • Resource-Constrained Deployment: Practical for deployment on GPUs with 32GB or more, or even 24GB for shorter contexts, due to its efficient MoE architecture and optimized KV cache.

Limitations

  • As a distillation, it does not aim to surpass its teachers in pure reasoning.
  • Training data for agentic tasks skews towards web/app/game/web3 coding, potentially limiting performance on out-of-distribution agent tasks.
  • Tool-name binding is loose and may require downstream normalization.