abacusai/Smaug-Mini

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Smaug-Mini is a 27 billion parameter agentic fine-tune of the Qwen3.8-27B model, developed by Abacus.AI. It features a dense hybrid-attention transformer architecture with a vision tower and a 262,144-token context length. This model is specifically optimized for end-to-end agentic tool use and automation, demonstrating significant performance improvements on benchmarks like AutomationBench and JobBench. While maintaining the base model's general capabilities, Smaug-Mini excels in complex multi-turn tasks requiring tool interaction and automation.

Loading preview...

Smaug-Mini: Agentic Fine-tune for Enhanced Automation

Smaug-Mini, developed by Abacus.AI, is a 27 billion parameter model fine-tuned from Qwen3.8-27B. It retains the base model's dense hybrid-attention transformer architecture, vision tower, and an extensive 262,144-token context length. The primary focus of this fine-tune is to significantly improve end-to-end agentic tool use and automation performance.

Key Capabilities & Performance

Smaug-Mini demonstrates substantial gains in agentic benchmarks:

  • AutomationBench: +4.5 points over its base (41.8 vs 37.3)
  • JobBench: +17.1 points over its base (50.5 vs 33.4)
  • LiveBench overall: +2.0 points (76.9 vs 75.3)
  • IFBench: +2.5 points (82.0 vs 79.5)

These improvements are achieved through on-policy reinforcement learning (GRPO) using multi-turn, tool-using automation episodes with outcome-based rewards. Notably, Smaug-Mini redistributes deliberation, leading to more reasoning per action and fewer actions per episode, completing tasks about three steps sooner. It also drastically reduces failure rates, with episodes burning the entire step budget dropping from 3.4% to 1.0%, without compromising general capabilities like reasoning and knowledge (GPQA-diamond scores remain at parity).

Use Cases & Deployment

Smaug-Mini is ideal for applications requiring robust agentic behavior, complex automation, and multi-turn tool interaction. It serves as a direct drop-in replacement for Qwen3.8-27B, supporting the same architecture and vision capabilities. Recommended sampling parameters for agentic use include temperature 1.0, top_p 0.95, and xhigh reasoning effort. The model's training involved a LoRA adapter over the language trunk, merged as a full delta, while the vision tower remains identical to the base model.

Finetunes
1 models