Abacus.AI introduced the Smaug line of open-weight models for enterprise agentic AI, the company said in an August 2026 platform update dated 1 August 2026, with product detail on its Smaug page.
Three models ship in the line: Smaug Flash, based on DeepSeek V4 Flash 0731, aimed at always-on self-improving enterprise agents; Smaug Mini, based on Qwen3.8 27B, for multimodal and smaller reasoning workloads; and Smaug Agentic, a Kimi K3 mixture-of-experts fine-tune at frontier scale. Training combines human-curated real-world agentic traces with synthetic hard examples, targeting agentic coding, tool use, automation and long-context reasoning. Weights are distributed via RouteLLM API and Hugging Face.
Abacus reported LiveBench overall of 77.4 for Smaug Flash versus 74.2 for the base, and agentic coding of 61.1 versus 46.8 on its harness (LiveBench dated 25 June 2026). For Smaug Agentic it cited GPQA Diamond 94.1 versus 93.5, DeepSWE 69.9 versus 67.5, LiveBench Agentic Coding 64.6 versus 62.2, and a median of 78 agent steps across 113 coding tasks over more than seven hours of continuous work. Benchmark scores are Abacus evaluation claims and have not been independently reproduced by AI-360. The launch is distinct from the company's 2021 Series C and computer-vision-as-a-service coverage already on AI-360.