XNews

Kimi K3 Release and Performance Benchmarks

Kimi K3 Release and Performance Benchmarks

Moonshot AI released Kimi K3, a 2.8-trillion-parameter native multimodal model with 1-million-token context window. It introduces Kimi Delta Attention for up to 6.3x faster decoding in long contexts and Attention Residuals for approximately 25% higher training efficiency at under 2% additional cost. On the DeepSWE software engineering benchmark, Kimi K3 max scored pass@1 68.5% versus GPT-5.6 Sol max's 72.7%, but surpassed it at pass@2 (82.0% vs 81.0%) and pass@4 (89.4% vs 85.8%). Cost per rollout is $4.65 for Kimi versus $8.37 for Sol; per $100, Kimi solves about 14.7 tasks versus Sol's 5.3. Median latency is 66 minutes for Kimi versus 17 minutes for Sol, with Kimi taking more steps. Language strengths differ: Kimi stronger in Rust, Sol in Python/TypeScript/JavaScript, tied on Go. Domain strengths split with Sol winning 5 of 8 categories. Task-level correlation is low (0.46), and the two-model union covers 108 of 113 tasks (95.6%). A recommended cascade strategy—Kimi first, then Sol on failure—achieves 85.6% solve rate at $7.30 per task. Open weights are planned by July 27, 2026.

Key insights

Kimi K3 Competes with GPT-5.6 Sol on Software Engineering Through Multiple Attempts

While trailing on pass@1, Kimi K3 overtakes Sol at pass@2 and pass@4, indicating strong iterative improvement capability. The low correlation between models' success patterns makes them highly complementary, with a combined coverage of 95.6% of benchmark tasks.

Cost Advantage Offset by Higher Latency

Kimi K3 costs roughly half per rollout compared to Sol, but median latency is nearly 4x longer (66 vs 17 minutes). The higher step count suggests different reasoning behavior, and inference optimization could narrow the latency gap.

Architectural Innovations Target Long-Context Efficiency

Delta Attention provides up to 6.3x faster decoding in million-token contexts, and Attention Residuals deliver 25% higher training efficiency at minimal cost. These innovations directly address the computational bottlenecks of long-horizon agentic coding.

Source posts