Abacus.AI ships open-weight Smaug models it says close the gap to frontier agents at a fraction of the cost
Smaug Agentic, Flash and Mini are open-weight models fine-tuned for long-running agent loops, pitched as a 10-100x cheaper substitute for Opus- and GPT-class models in enterprise agentic workflows.
Abacus.AI on September 10 released Smaug, three open-weight models fine-tuned for long-running, self-improving agent loops over single-turn chat. Smaug Agentic is a 2 trillion-parameter model built on Kimi K3 for complex coding loops. Smaug Flash, built on DeepSeek Flash, targets persistent personal agents that connect to WhatsApp, Telegram and Slack. Smaug Mini is a 27-billion-parameter model for multimodal and smaller reasoning workloads. All three are downloadable on Hugging Face and available through Abacus.AI's RouteLLM API.
The pitch is cost
CEO Bindu Reddy concedes the gap: open-weight models, she said, are "rapidly closing the gap to frontier closed models, but still underperform in long-running agent loops." Smaug is pitched as the fix at a fraction of the price. Abacus.AI says its fine-tuning technique, which combines human-curated agentic traces with synthetic data drawn from difficult examples, lifts long-running agentic-loop performance 15-20% without added inference cost. It also says Smaug Agentic ran 113 DeepSWE coding tasks over more than seven hours with zero infrastructure errors or timeouts, on the company's own benchmark. Enterprises can run any of the three inside their own VPC or on in-house GPU clusters.
This launch is single-sourced to Abacus.AI's announcement. As of publication no independent outlet with a confirmed byline had reviewed the benchmark claims, so treat the performance numbers as vendor-reported. If the cost case interests you, run Smaug against your current model on your own agent tasks and compare measured outcomes before switching.