Abacus.AI’s Smaug Agentic fine-tune targets the failure tail in long-running coding agents

On September 10, Abacus.AI released Smaug Agentic alongside Smaug Flash and Smaug Mini. Smaug Agentic is a supervised agentic fine-tune of Moonshot AI’s Kimi K3 with the same architecture, tokenizer, 1M-token context and vision encoder as the base model. Abacus says it trained the model on filtered multi-turn, tool-using coding trajectories and published the weights on Hugging Face under the inherited Kimi K3 licence. The model card reports modest benchmark gains over the base on several coding and reasoning tasks, but the more interesting claim is behavioural: Abacus says 99th-percentile reasoning length falls to roughly 0.6× of Kimi K3 on two unrelated benchmarks while visible answer length remains statistically unchanged.

Long-running coding agents often fail economically before they fail functionally: an agent can keep reasoning, calling tools and consuming tokens while making little progress. Smaug Agentic is a concrete attempt to improve that failure tail through post-training rather than a new foundation model or a bigger inference budget. Because the architecture is unchanged, teams already able to serve Kimi K3 can evaluate Smaug as a drop-in alternative using the same inference stack. If the reported behaviour holds on real repositories, post-training aimed at agent-loop stability could become as important as single-task benchmark scores when builders choose open models for autonomous coding.

Smaug Agentic keeps Kimi K3’s underlying 2.8T-parameter mixture-of-experts architecture, 104B activated parameters, 1,048,576-token context and MoonViT-V2 vision encoder. Abacus says only behaviour was adapted through supervised fine-tuning on curated multi-turn, tool-using coding trajectories. That means the deployment proposition is unusually simple for an experimental agent model: infrastructure that already serves Kimi K3 should be able to serve Smaug Agentic without a new model architecture.

Abacus reports that normal deliberation remains broadly unchanged while the long tail of reasoning shrinks sharply. Its model card says p99 reasoning length falls to about 0.6× of the base model on SciCode and about 0.55× on a long-context reasoning test, while visible answer length stays statistically similar. Across 113 DeepSWE tasks and more than seven hours of continuous work, Abacus reports no infrastructure errors or timeouts. These are vendor-run results, but they target a production concern that ordinary benchmark tables often hide: whether an agent keeps burning time and tokens after it has stopped making useful progress.

In Abacus’s published runs, Smaug Agentic scores 69.9 on DeepSWE versus 67.5 for Kimi K3, 64.6 versus 62.2 on LiveBench Agentic Coding and 60.8 versus 58.7 on SciCode. It performs worse than the base on Terminal-Bench 2.1 in the published table. Abacus also notes that some comparison columns reproduce other vendors’ published numbers rather than rerunning every model under one identical harness. Builders should therefore treat the release as a testable candidate rather than a settled ranking.

The checkpoint is enormous: Kimi K3-class serving remains a server-scale deployment even though the weights are downloadable. Abacus documents vLLM and SGLang serving paths and says the model works with OpenAI-compatible agent frameworks, but self-hosting a 2.8T MoE still requires substantial accelerator infrastructure. The operational choice is therefore not simply free open model versus paid API; it is control and potentially lower marginal inference cost in exchange for significant serving complexity.