Abacus.AI's Smaug line pitches open-weight agent models at…

Three fine-tuned open-weight models — Smaug Agentic, Flash, and Mini — target long-running agent loops, with the company claiming 15–20% performance gains over their base models.

Abacus.AI released the Smaug line on Sept. 10, 2026, a family of three open-weight language models the San Francisco company says deliver a 15–20% performance lift over their base models on agentic workloads while running at 10 to 100 times lower cost than frontier APIs from Anthropic and OpenAI. The pitch is aimed squarely at teams running long, tool-calling agent loops, where token bills compound by the hour.

The lineup fine-tunes three well-known open-weight bases. Smaug Agentic is built on Moonshot AI’s Kimi K3, a mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated, a 1,048,576-token context window, and the MoonViT-V2 vision encoder. Smaug Flash sits on DeepSeek V4 Flash 0731 and preserves the 1M-token context, with three LoRA adapters merged as full deltas. Smaug Mini is a 27-billion-parameter fine-tune of Qwen3.8 27B. All three are downloadable on Hugging Face and callable through Abacus.AI’s RouteLLM API.

The benchmark table does most of the talking. Smaug Flash posts 61.1 on LiveBench agentic coding against the base’s 46.8, 38.83 on AutomationBench strict-pass 600 against 25.1, and 73.3 on NL2repo-bench against 54.2. Its LiveBench overall score of 77.4 edges past the 76.0 Abacus cites for Claude Sonnet 5, using the June 25, 2026 leaderboard. Smaug Agentic notches 94.1 on GPQA Diamond, 69.9 on DeepSWE, 86.5 on Terminal-Bench 2.1, and 81.0 on MMMU-Pro. Mini’s JobBench score rises to 50.5 from a base of 33.4.

The more revealing numbers are operational. On an 8×B300 deployment at temperature 1.0 with reasoning effort at max, Smaug Agentic ran 113 DeepSWE tasks over more than seven hours, median 78 agent steps per task, with zero infrastructure errors and zero timeouts. P99 reasoning length came in at roughly 0.6× the base on SciCode and AA-LCR. That’s the shape of a model tuned for the tedious middle of an agent’s day, not for a leaderboard sprint.

CEO Bindu Reddy and co-founder Arvind Sundararajan, both Google, Amazon and Uber veterans, have built Abacus.AI into a platform they say serves over 3 million professionals, small businesses and enterprises, backed by Coatue, Tiger Global and Index Ventures. The weights ship under the Kimi K3 License, servable on vLLM, SGLang and TokenSpeed. Whether the 15–20% lift holds outside Abacus’s own evaluations is the question every buyer will ask; TechEdgeAI flagged that the performance claim comes from the company and should be distinguished from independent benchmark validation. The cost delta, if it survives contact with production, is the part that reorders the market.

Abacus.AI's Smaug line pitches open-weight agent models at 10–100x lower cost than frontier APIs

Sources