San Francisco – September 15, 2026 -- Abacus.AI has released a new family of open-weight large language models, Smaug, claiming 10-100x lower operating costs than frontier models from Anthropic and OpenAI while improving long-running agentic loop performance by 15-20%.
Three Models Target Different Enterprise Agentic Workloads
The Smaug line comprises three models: Smaug Agentic, Smaug Flash, and Smaug Mini, all released as open-weight and downloadable via Hugging Face. Smaug Agentic is a 2-trillion-parameter model built on Kimi K3, positioned as a cost-efficient replacement for Opus-class models in complex coding loops, with the option for enterprises to host it on in-house GPU clusters for security and data privacy. Smaug Flash, fine-tuned on DeepSeek Flash, is optimized for personal agents that connect to WhatsApp, Telegram, and Slack, maintaining long-running conversations that adapt to users over time. Smaug Mini is a compact 27B model designed for multimodal use cases, smaller reasoning workloads, and enterprise chatbots, and can be further fine-tuned on a company's own data.
Fine-Tuning Technique Applies Across Any Open-Source Base Model
Abacus.AI describes Smaug as a novel fine-tuning method that can be applied to any open-source base model to boost agentic loop performance without added compute cost. The company frames this as enabling self-improving enterprise AI agents at scale while enterprises retain full control over data by hosting models within their own cloud VPC environment.
Benchmark Results Published on LiveBench AI
Abacus.AI has published comparative results against base models on LiveBench AI, covering categories including agentic coding, tool use, automation, and long-context reasoning. The company states the Smaug models deliver measurable gains over their respective base models across these workload categories.
CEO Frames Release as Evidence Open-Weight Models Can Match Frontier Systems
Bindu Reddy, CEO of Abacus.AI, said open-weight models are closing the performance gap with closed frontier models but still lag in long-running agent loops, a shortfall the Smaug line is designed to address while maintaining a 10-100x cost advantage over closed-source alternatives. Abacus.AI positions the release as both a commercial product and a demonstration that targeted fine-tuning methodology can allow open-weight models to compete with, and in some cases surpass, frontier models on agentic AI tasks.
Company Plans to Expand the Line as Base Models and Agentic Data Grow
Abacus.AI said it expects to continue advancing the Smaug line as underlying open-source base models evolve and as its library of agentic traces expands, without specifying a release timeline for future versions.