Abacus.AI released three open-weight Smaug models on Hugging Face, with Smaug Flash posting the biggest gains: +14.3 points in LiveBench agentic coding, +13.7 in AutomationBench and +19 in NL2Repo-Bench.
Those improvements target a core weakness in base models that lose context or spin in long tool-use loops; Abacus says it used the same fine-tuning recipe across three different bases to strengthen real multi-step agent behavior.
Smaug Mini, a 28B model, beat Claude Sonnet 5, Claude Opus 4.6 and GPT-5.6 Luna on IFBench, AutomationBench and JobBench, even as its LiveBench agentic coding score slipped 0.6 points from its base.
Smaug Agentic showed only a +0.3 overall benchmark gain, but completed 113 tasks over 7+ hours with a median 78 agent steps and zero infrastructure errors, underscoring reliability over headline scores.
The results suggest the same training method helps most where each base model is weakest, making workload fit—not a single benchmark—the key test for developers choosing agentic AI models.