LiamVisionary/swarm-sovereign-scout-12b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

LiamVisionary/swarm-sovereign-scout-12b is a 12 billion parameter Gemma 4-based fine-tuned model, specialized for low-refusal coding, terminal workflows, and agentic tasks with a 32768 token context length. Derived from yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2, it excels at reasoning-heavy technical work and native Gemma 4 tool-use. This model is optimized for autonomous agent loops where reduced refusal behavior is critical.

Loading preview...

Swarm Scout 12B: Low-Refusal Agentic Coding Model

Swarm Scout 12B, developed by LiamVisionary, is a 12 billion parameter fine-tuned model based on google/gemma-4-12B-it and further refined from yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2. It features a 32768 token context length and is specifically designed for agentic workflows, emphasizing reduced refusal behavior through an "abliteration" pass.

Key Capabilities

  • Coding and Debugging: Provides assistance for programming tasks.
  • Terminal-Style Agent Workflows: Supports read, reason, act, and verify loops for technical tasks.
  • Native Gemma 4 Tool-Use: Implements the native tool-use protocol for Gemma 4 models.
  • Reasoning-Heavy Technical Work: Excels in complex problem-solving with configurable thinking.
  • Low-Refusal Behavior: Engineered to minimize refusal responses, making it suitable for autonomous agent loops.
  • Multimodal Support: Includes verified image input capabilities via a companion GGUF projector artifact.

Good For

  • Developers building autonomous agents requiring reliable, low-refusal responses.
  • Applications involving coding assistance, debugging, and terminal-based automation.
  • Technical tasks that benefit from strong reasoning and multi-step agentic execution.
  • Use cases leveraging native Gemma 4 tool-use protocols.

Limitations

  • Specialized for coding and agentic work; general knowledge answers may require external verification.
  • Primarily English-centric.
  • Low-refusal behavior is intentional and requires application-level guardrails for safety and abuse prevention.

Finetunes
2 models