abacusai / Smaug-Agentic

---

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 continues the Smaug line from Smaug-72B-v0.1 — the first open-source model to pass an average score of 80% on the HuggingFace Open LLM Leaderboard — scaled to a frontier MoE base and targeted at long-horizon tool use and coding agents.

The fine-tune adapts behaviour only. Every architectural parameter is unchanged from Kimi K3, so any inference stack that serves K3 serves this model.

Key Features

• Agentic Post-Training : Supervised fine-tuning on filtered multi-turn, tool-using coding trajectories, with reasoning tokens masked from the loss — the model sees its own thinking in context but is never supervised on it.
• Architecturally Identical to Kimi K3 : Same 2.8T MoE, same 1M-token context, same MoonViT-V2 vision encoder, same tokenizer. A drop-in replacement for existing K3 deployments.
• Bounded Deliberation : Typical reasoning length is unchanged while the runaway tail collapses — p99 reasoning falls to 0.62× of base on SciCode and 0.55× on AA-LCR, with no loss of answer detail.
• Interleaved Thinking Preserved : reasoning_content survives across turns, so multi-step agent loops keep the model's chain of thought.
• Native MXFP4 Weights : Inherited from the base model, for broad hardware compatibility.

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.

Notes

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. Behavioral Differences vs. Kimi K3

The benchmark deltas above are small. The behavioural change is not. Everything below is measured on paired runs at identical sampling settings (temperature = 1.0, top-p = 0.95, reasoning effort 'max', same completion cap, same task set), so none of it is a temperature artifact.

4.1 Bounded deliberation — typical reasoning is unchanged, the runaway tail collapses

The single largest behavioural difference. Per-subproblem paired comparison on SciCode (288 subproblems, identical protocol), measuring characters of reasoning_content :

The median is untouched; the tail is halved. Counting the pathological cases directly:

This replicates on a completely different benchmark. On AA-LCR (100 long-context questions, ~95k-token prompts) median reasoning is actually slightly longer for Smaug-Agentic (2,573 vs 2,308 chars), while p99 drops from 39,945 to 21,836 ( 0.55× ) and the maximum from 40,344 to 30,734.

Two unrelated task distributions, same signature: normal deliberation is preserved, unbounded deliberation is suppressed. In practice this shows up as fewer requests that burn the entire token budget without producing an answer.

4.2 It is not simply terser

A natural worry is that the fine-tune just shortened everything. It did not. On GPQA Diamond, where both arms ran on the same serving stack, the visible answer length distribution is statistically indistinguishable:

Smaug-Agentic is shorter on 103 of 198 questions — a coin flip. Answer style is inherited from the base model intact; what changed is the willingness to deliberate without bound.

4.3 One decisive shell action per turn, sustained over very long loops

This is what the training data optimises for, and it transfers cleanly. Across 113 DeepSWE tasks and 7h33m of continuous agentic work: 0 infrastructure errors, 0 timeouts , and a median of 78 agent steps per task (p90 159, max 218). Tasks that passed used more steps than tasks that failed (median 80 vs 70), i.e. failures are not the model giving up early.

5. Deployment

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:

• vLLM — see recipes
• SGLang — see cookbook
• TokenSpeed — see recipes

---

6. 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.

---

7. 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.

---

8. 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}
}

---

9. Contact Us

If you have any questions, please reach out at Abacus.AI .