Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost
Cognition released SWE-2, a coding model post-trained via RL from Moonshot AI's 2.8T-parameter Kimi K3, achieving 50.0% on FrontierCode 1.1 Main—within 1 point of Fable 5.1 at 64% lower cost SWE-2 is the first Cognition model with selectable reasoning-effort levels trained in a single RL run using Pareto-informed linear cost penalties, reducing turns by 58% and cost by 81% versus SWE-1.7 The model demonstrates significant RL headroom on top of the massive K3 base, adding 5–6 points on most bench
Analysis
TL;DR
- Cognition released SWE-2, a coding model post-trained via RL from Moonshot AI's 2.8T-parameter Kimi K3, achieving 50.0% on FrontierCode 1.1 Main—within 1 point of Fable 5.1 at 64% lower cost
- SWE-2 is the first Cognition model with selectable reasoning-effort levels trained in a single RL run using Pareto-informed linear cost penalties, reducing turns by 58% and cost by 81% versus SWE-1.7
- The model demonstrates significant RL headroom on top of the massive K3 base, adding 5–6 points on most benchmarks while showing focused exploration behavior with fewer detours
- SWE-2 remains proprietary with no open weights or standalone API, accessible only through Devin Desktop, CLI, and rolling out to Web and Fusion
- Despite strong performance on Terminal-Bench 2.1 (92.8%) and FrontierCode, SWE-2 shows a clear weakness on Terminal-Bench 4 (27.3%), trailing top competitors by ~30 points
Why It Matters
SWE-2 demonstrates that scaling reinforcement learning to multi-trillion-parameter models yields substantial performance gains, proving that even state-of-the-art bases like Kimi K3 have significant untapped potential through task-specific RL post-training. For AI practitioners, it highlights the emerging pattern of "effort-level" models that let users trade compute for performance, and the importance of specialized coding benchmarks in evaluating real-world agent capability beyond generic LLM scores.
Technical Details
- Training pipeline: Post-trained from Kimi K3 (2.8T-parameter MoE) using a novel single-run RL algorithm that jointly trains three effort levels (medium/high/max) with slope-matched linear cost penalties, where the reward function is R = S − λC (binary success minus cost-weighted penalty)
- Cost-performance frontier: Each effort level's λ is set to the local slope of the base model's Pareto curve, making the iso-reward line tangent to the frontier so reward can only improve by pushing the frontier upward—not just shifting along it
- Efficiency gains: Length-weighted reward baseline (sum(R×L)/sum(L)) stabilizes training without extra compute; DSpark speculative decoding with a SpecForge-retrained draft model achieves 15% longer accept lengths; NVFP4 and FP8 kernels with quantization-aware training reduce train-inference mismatch
- Benchmark performance: FrontierCode 1.1 Main 50.0%, DeepSWE 1.1 73.0%, Terminal-Bench 2.1 92.8%, Terminal-Bench 4 27.3%; leads on Terminal-Bench 2.1 and beats Kimi K3 base by 5–6 points across most metrics
- Behavioral improvements: SWE-2 medium makes its first real edit after a median of 18 steps versus 48 for SWE-1.7; shows stronger end-to-end test coverage, resourcefulness when tools are blocked, and verification discipline (re-deriving rather than re-asserting conclusions)
Industry Insight
The "effort-level" paradigm—where a single model serves multiple reasoning budgets trained simultaneously—is likely to become a standard pattern for production AI agents, allowing teams to dynamically allocate compute based on task complexity rather than maintaining separate model variants. The proprietary-only distribution strategy (Devin exclusively) signals that top-tier coding models are increasingly treated as competitive moats rather than open commodities, reinforcing the trend where value shifts from raw model weights to the agent infrastructure and RL recipes built around them. Finally, the significant gap between Terminal-Bench 2.1 (92.8%) and Terminal-Bench 4 (27.3%) suggests that current benchmarks still fail to capture consistent long-horizon reasoning ability, prompting the industry to develop more rigorous multi-step evaluation suites.
Disclaimer: The above content is generated by AI and is for reference only.