Abacus AI Ships Smaug-Agentic: Kimi K3 Fine-Tune Beats Fable 5 on Four of Five Agentic Coding Benchmarks
Abacus AI released Smaug-Agentic on August 11, a fine-tuned version of Kimi K3 built specifically for agentic task completion. The model lands at number 5 on LiveBench overall with a score of 79.5 and outperforms Claude Fable 5 Max Effort on four of five published agentic coding benchmarks.
The model is available on HuggingFace and costs $0.329 per successful task on LiveBench, compared to $1.439 for Fable 5 Max Effort and $0.348 for stock Kimi K3.
Benchmark Numbers
Abacus AI’s published results compare Smaug-Agentic directly against Kimi K3 and Claude Fable 5:
| Benchmark | Smaug-Agentic | Kimi K3 | Fable 5 |
|---|---|---|---|
| LiveBench Agentic Coding | 64.6 | 62.2 | 62.2 |
| DeepSWE | 69.9 | 67.5 | 70.0 |
| SciCode | 60.8 | 58.7 | 60.2 |
| AutomationBench | 31.0 | 30.8 | 29.1 |
| Terminal-Bench 2.1 | 86.5 | 88.3 | 88.0 |
On four dimensions it beats or ties both — the exception is Terminal-Bench 2.1, where it trails Kimi K3 by 1.8 points and Fable 5 by 1.5.
On LiveBench overall, it scores 79.5 against Kimi K3’s 79.2. The gap is narrow, but the distribution matters: Smaug-Agentic improved specifically where Abacus targeted, with reasoning held flat and agentic coding pulled forward.
The Training Recipe
The lead researcher at Abacus AI described the core problem the training was designed to solve: vanilla fine-tuning at this scale tends to improve format compliance but degrades reasoning. Their approach combined reinforcement learning, supervised fine-tuning, and preference optimization into a single pass, with the explicit constraint of preserving Kimi K3’s base reasoning capabilities.
The result is a model that improves on agentic tasks — coding agents, tool use, multi-step automation — without the typical regression on the benchmarks the base model was strong on. Reasoning scores on LiveBench are essentially flat between Smaug-Agentic (90.3) and Kimi K3 (90.7).
What This Means
Kimi K3 entered GitHub Copilot this week and is being adopted at enterprise scale at $3/$15 per million tokens. Smaug-Agentic is positioned as a drop-in replacement — same license model (Kimi K3 license), same model family, improved agentic behavior.
The model is not from a frontier lab and does not introduce new architecture. What it demonstrates is that post-training RL can move an already-strong open-weight model meaningfully on agentic tasks without paying the reasoning-regression tax that has limited many previous fine-tuning efforts at this scale.
The Terminal-Bench score is the one number worth tracking: at 86.5 it still beats most frontier models on that benchmark, but it’s the one category where Kimi K3 pulls ahead. Whether that gap closes in a subsequent release will determine if Smaug-Agentic holds the open-weight agentic coding lead.
Key Details
- Model: Smaug-Agentic (Kimi K3 base, fine-tuned by Abacus AI)
- Access: HuggingFace (open weight)
- Cost: $0.329 per successful LiveBench task
- LiveBench Overall: 79.5 (#5 globally as of August 12)
- Training: RL + SFT + preference optimization on Kimi K3
- License: Inherits Kimi K3 license (MIT-like, with revenue gate)
- Planned integration: Abacus AI’s own coding agent platform