Abacus.AI on September 10, 2026, introduced the Smaug line, three open-weight language models fine-tuned for enterprise agentic workloads: Smaug Agentic, Smaug Flash, and Smaug Mini. All three are available for download on Hugging Face and can also be used through the company’s RouteLLM API.
The models were introduced across Abacus.AI’s enterprise agentic AI platform and Super Assistant. The company said enterprises can host the models inside their own cloud VPC environment, giving them full control over their data and where their AI is hosted, and that organizations concerned about security and data privacy can host Smaug Agentic on an in-house GPU cluster.
Abacus.AI describes Smaug as a fine-tuning technique applicable to any open-source base model, and said it improves the performance of long-running agentic loops by 15–20% without increasing cost. The company said this enables enterprises to deploy self-improving AI agents at scale at open-source model costs, which it described as typically 10–100× lower than those of frontier models from Anthropic and OpenAI.
“Open-weight models are rapidly closing the gap to frontier closed models, but still underperform in long-running agent loops,” Bindu Reddy, CEO of Abacus.AI, said in the company’s announcement. Reddy said the Smaug line addresses that shortcoming while remaining 10–100× cheaper than closed-source models.
According to the company’s technical brief, the line grew out of the cost and inefficiency of running large agentic loops with large contexts and repeated tool calls, compounded by prompt caching breaking down across time gaps. The methodology combines human-curated, real-world agentic traces with synthetic data grounded in hard examples, and the brief reports consistent lifts across agentic coding, real-world tool use, automation, and long-context reasoning and instruction following when applied to a range of open-weight base models.
Smaug Agentic
Smaug Agentic, the largest model in the line, is a supervised fine-tune of Moonshot AI’s Kimi K3, a mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated parameters, a 1,048,576-token context length, and the MoonViT-V2 vision encoder, according to its model card. The fine-tune changes no architectural parameters, and the model is released under the Kimi K3 License inherited from the base model, whose terms users must follow. The card says training used curated multi-turn, tool-using coding trajectories with reasoning tokens masked from the loss, so that training steers the model’s actions while leaving the base model’s reasoning distribution intact; dataset contents are not disclosed.
The card reports scores of 94.1 on GPQA Diamond, 75.7 on AA-LCR, 69.9 on DeepSWE, 86.5 on Terminal-Bench 2.1, 60.8 on SciCode, 64.6 on LiveBench agentic coding, 31.0 on AutomationBench, and 81.0 on MMMU-Pro. It states gains over the Kimi K3 base of 2.4 points on DeepSWE, 2.4 on LiveBench agentic coding, 2.1 on SciCode, 1.0 on AA-LCR, and 0.6 on GPQA Diamond.
The card also reports that p99 reasoning length falls to roughly 0.6× of the base model on SciCode and AA-LCR, while visible answer length remains statistically indistinguishable from Kimi K3. Across 113 DeepSWE tasks and over seven hours of continuous work, the company said it recorded zero infrastructure errors and zero timeouts. Because the architecture is unchanged, Smaug Agentic runs anywhere Kimi K3 runs, with published serving recipes for vLLM, SGLang, and TokenSpeed. The announcement positions the model as a cost-efficient replacement for Opus-class models.
Smaug Flash and Smaug Mini
Smaug Flash is fine-tuned on DeepSeek V4 Flash 0731 and targets enterprise self-improving agents. The brief says the base model is susceptible to “spins and confusion” in long-context agentic tool use, and that Smaug Flash addresses this while maintaining the base model’s cost and speed advantages. The announcement describes Smaug Flash as optimized for personal agents that can connect to messaging apps including WhatsApp, Telegram, and Slack and maintain long-running conversations with users. The brief’s reported scores against the DeepSeek V4 Flash 0731 base, with Claude Sonnet 5 listed as a reference column, include 77.4 versus 74.2 on LiveBench overall, 61.1 versus 46.8 on LiveBench agentic coding, 56.6 versus 54.4 on DeepSWE 1.1, 38.83 versus 25.1 on AutomationBench’s public 600 under strict pass, and 73.3 versus 54.2 on NL2repo-bench.
Smaug Mini is a 27-billion-parameter model fine-tuned on Qwen3.8 27B, aimed at multimodal use cases and smaller reasoning workloads. The brief reports 76.9 versus the base model’s 75.3 on LiveBench overall, 82.0 on IFBench, 41.8 versus 37.3 on AutomationBench, 50.5 versus 33.4 on JobBench’s official 65-task protocol, and 55.8 versus 42.3 on NL2repo-bench, with Claude Sonnet 5, Claude Opus 4.6, and GPT-5.6 Luna listed as reference columns. The announcement describes Smaug Mini as suited to simple tasks and enterprise chatbots, and says enterprises can further fine-tune it on their own data.
Evaluation Protocol and Roadmap
The technical brief states that evaluations were run by Abacus.AI under the same harness for every model unless marked with a dagger as a reported score not from its harness, and that LiveBench rows are drawn from the published LiveBench leaderboard dated June 25, 2026. The Smaug Agentic card says its results were produced on a dedicated 8×B300 deployment at a temperature of 1.0 with reasoning effort set to max, and that its SciCode evaluation includes a repair for an upstream defect that left 12 subproblems unwinnable. The company said the models are listed on LiveBench under the leaderboard’s finetunes filter.
Abacus.AI said the Smaug line is both a product release and a demonstration of its thesis that, with the right fine-tuning methodology, open-weight models can compete with and surpass frontier models on agentic AI tasks. The company said it expects to continue advancing the line as base models evolve and its library of agentic traces grows.
Credit: Source link
