1. Model Introduction
Smaug-Agentic is an agentic fine-tune of Kimi K3, the 2.8T-parameter Mixture-of-Experts model from Moonshot AI, finetuned by Abacus.AI. It improves on the base across reasoning and agentic benchmarks (+2.4 on DeepSWE, +2.4 on LiveBench agentic-coding, +2.1 on SciCode, +1.0 on AA-LCR, +0.6 on GPQA Diamond) and matches or leads every compared frontier model on GPQA Diamond, AA-LCR, SciCode, LiveBench agentic coding, and AutomationBench, while holding overall capability at parity (+0.3 overall LiveBench).
The fine-tune adapts behaviour only. Every architectural parameter is unchanged from Kimi K3 — same 2.8T MoE, same 1M-token context, same MoonViT-V2 vision encoder, same tokenizer — so any inference stack that serves K3 serves this model as a drop-in replacement.
This card describes the training approach and the evaluation results. Dataset contents are not disclosed; training data consists of filtered multi-turn, tool-using coding trajectories.
2. Model Summary
3. Evaluation Results
This table lists only the benchmarks we ran ourselves. Reference columns reproduce the officially published Kimi K3 numbers and the accompanying figures for the other models; we did not re-run them.
LiveBench category profile
Smaug-Agentic (Kimi K3 finetune) against its base and Claude Fable 5. Scores 0–100 from the LiveBench 2026-06-25 public leaderboard at max effort; overall = mean of the seven category averages.
All Smaug-Agentic results were produced on a dedicated 8×B300 deployment at temperature = 1.0 and reasoning effort 'max', following the Kimi K3 top-p convention: top-p = 0.95 for single-step tasks, top-p = 1.0 for agentic tasks.
- GPQA Diamond and AA-LCR scores are the average of 3 runs.
- DeepSWE: top run using mini-swe-agent.
- Terminal-Bench 2.1: scored with the Terminus 2 agent; the official Kimi K3 number uses the Kimi Code agent, on which we scored lower (76.4).
- SciCode: test split, prompt with background. Includes a repair for an upstream gold-injection defect that leaves 12 subproblems unwinnable; contributed upstream as scicode-bench/SciCode#61.
- LiveBench: complete data for all models on the LiveBench leaderboard.
- MMMU-Pro:
standard(10-option) setting, single pass, no tool augmentation.
4. Training Approach
Smaug-Agentic was trained to keep long-horizon agentic coding loops decisive and free of runaway deliberation at max reasoning effort. It is a supervised fine-tune on curated multi-turn, tool-using coding trajectories: the model sees its own reasoning in context, but reasoning tokens are masked from the loss, so the base model's reasoning distribution is preserved while its actions are steered. No architectural changes were made.
5. Known Behaviors and Limitations
Typical deliberation is unchanged while the runaway tail collapses: p99 reasoning length falls to roughly 0.6× of the base on two unrelated benchmarks (SciCode and AA-LCR), which in practice means fewer requests that burn the entire token budget without producing an answer. Answers are not terser — visible answer length is statistically indistinguishable from Kimi K3. Long agentic loops stay stable: across 113 DeepSWE tasks and over seven hours of continuous work we saw zero infrastructure errors and zero timeouts, and tasks that passed used more steps than tasks that failed — failures are not the model giving up early.
6. Deployment
[!Note] Because the architecture is unchanged, Smaug-Agentic runs anywhere Kimi K3 runs. The inference engines below serve it with the recipes published for the base model:
7. Model Usage
Sampling behaviour is inherited from Kimi K3, and these settings were used for every number above.
Smaug-Agentic always has thinking enabled, and will return reasoning_content. Thinking effort is configured with the top-level reasoning_effort request field, which supports "low", "high", and "max" (default "max"). Set temperature = 1.0, with top_p = 0.95 for single-step tasks and top_p = 1.0 for agentic tasks.
Like the base model, Smaug-Agentic was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, the complete assistant message returned by the API must be passed back to messages as-is — including reasoning_content and tool_calls, not just content.
Coding Agent Framework
The model is trained for multi-turn tool use and works with agent frameworks that speak the OpenAI chat-completions contract.
One practical note from our own evaluation: some OpenAI-compatible servers reject non-standard fields echoed back in conversation history (provider_specific_fields, function_call, annotations, refusal, audio). Strip those before re-sending. reasoning_content and tool_calls are accepted and should be kept, so that interleaved thinking survives across turns.
8. License
Smaug-Agentic is a derivative of Kimi K3 and is released under the Kimi K3 License, inherited from the base model. Users must comply with the base model's terms.
9. Citation
@misc{abacusai2026smaugagentic,
title = {Smaug-Agentic},
author = {Abacus.AI},
year = {2026},
note = {Agentic supervised fine-tune of moonshotai/Kimi-K3},
url = {https://huggingface.co/abacusai/Smaug-Agentic}
}
The Smaug line and the DPO-Positive method behind it:
@article{pal2024smaug,
title={Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive},
author={Pal, Arka and Karkhanis, Deep and Dooley, Samuel and
Roberts, Manley and Naidu, Siddartha and White, Colin},
journal={arXiv preprint arXiv:2402.13228},
year={2024}
}
10. Contact Us
If you have any questions, please reach out at Abacus.AI.