Kimi K3 Release and Performance

Moonshot AI released Kimi K3 on July 16, 2026, a 2.8-trillion-parameter native multimodal model with 1-million-token context, featuring Delta Attention for up to 6.3x faster decoding in long contexts and Attention Residuals for ~25% higher training efficiency. Open weights were promised by July 27, 2026. - In DeepSWE software engineering benchmarks against GPT-5.6 Sol max: pass@1 68.5% vs 72.7% (GPT), but pass@2 Kimi 82.0% vs 81.0%, pass@4 Kimi 89.4% vs 85.8%. - Kimi cost ~$4.65 per rollout vs Sol’s $8.37; per $100, Kimi solves ~14.7 tasks vs Sol’s ~5.3. - Latency: Kimi median ~66 minutes, Sol ~17 minutes. - Language strengths: Kimi stronger in Rust, Sol in Python/TS/JS; Go tied. - Domain splits: Sol leads in serialization, concurrency, program analysis; Kimi leads in ops tools, runtime internals; compliance near coin-flip. - Cascading strategy (Kimi first, then Sol on failure) reaches 85.6% solve rate at ~$7.30/task, beating Sol-only 72.3% at $8.37. Task correlation ~0.46 indicates complementary failure modes.
Key insights
Kimi K3 Offers Frontier Coding Performance at Half the Cost of GPT-5.6 Sol
On the DeepSWE software engineering benchmark, Kimi K3 achieved pass@1 68.5% vs GPT-5.6 Sol max 72.7%, but surpassed Sol on pass@2 (82.0% vs 81.0) and pass@4 (89.4% vs 85.8). Per-rollout cost was $4.65 for Kimi versus $8.37 for Sol, yielding ~2.8x more solved tasks per dollar. This cost advantage makes Kimi K3 a compelling choice for budget-sensitive coding workloads, especially when multiple retries are acceptable.
Complementary Strengths Enable Effective Cascading Strategies
Kimi K3 and GPT-5.6 Sol max exhibit low task-level correlation (~0.46), meaning they fail on different tasks. Kimi excels in Rust, ops tools, and runtime internals; Sol leads in Python/TypeScript, serialization, concurrency, and program analysis. A cascading setup—running Kimi first and escalating to Sol on failure—reached 85.6% solve rate at $7.30/task, beating Sol-only’s 72.3% at $8.37. This suggests that multi-model routing is more cost-effective than relying on a single frontier model.
Open Weights and Architectural Innovations Lower Barrier for Long-Context Deployment
Kimi K3 was announced with plans to open weights by July 27, 2026. Its Delta Attention mechanism achieves up to 6.3x faster decoding in million-token contexts, and Attention Residuals improve training efficiency by ~25% at under 2% additional cost. These innovations address the practical challenges of running 1M-context models at scale, making local and self-hosted deployment more feasible for organizations with data sovereignty requirements.