Key Highlights
- Kimi.ai launched Kimi K3 with a 2.8 trillion-parameter Mixture-of-Experts architecture.
- The model supports a 1 million-token context window and native multimodal capabilities.
- Kimi.ai claims Delta Attention enables up to 6.3x faster decoding for million-token contexts.
- Attention Residuals improve training efficiency by approximately 25% with less than 2% additional computational cost.
- Kimi K3 ranked first in Terminal Bench 2.1 with a score of 88.3.
- The model also led Program Bench with a score of 77.8 and BrowseComp with 91.2.
- Arena.ai ranked Kimi K3 No. 1 on the Frontend Code Arena leaderboard with 1,679 points.
- Open-weight versions of Kimi K3 are planned for release on July 27, 2026.
Kimi.ai has introduced Kimi K3, its latest frontier artificial intelligence model, highlighting significant advances in long-context reasoning, coding, multimodal understanding, and autonomous AI agent capabilities. According to the company, Kimi K3 is built using a 2.8 trillion-parameter Mixture-of-Experts (MoE) architecture and supports a context window of up to one million tokens.
The company said the model is designed to handle complex coding tasks, long-document reasoning, visual understanding, and self-evolving AI workflows. Kimi K3 is currently available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, while open-weight versions are scheduled for release on July 27, 2026.
Kimi.ai stated that Kimi K3 introduces Delta Attention, an optimization technique that enables up to 6.3 times faster decoding when processing million-token contexts. The company also said its Attention Residuals architecture improves training efficiency by around 25% while increasing computational cost by less than 2%.
Strong Coding Benchmark Results
Benchmark charts released by Kimi.ai show that Kimi K3 achieved competitive results across several coding evaluations.
In Terminal Bench 2.1, Kimi K3 scored 88.3, ahead of OpenAI’s GPT-5.6 Sol (85.8), Opus-4.8 (84.6), Fable 5 (84.6), GPT-5.5 (83.4), and GLM-5.2 (82.7).
The company also reported that Kimi K3 ranked first in Program Bench with a score of 77.8, narrowly outperforming GPT-5.6 Sol (77.6), followed by Fable 5 (76.8), Opus-4.8 (71.9), GPT-5.5 (70.8), and GLM-5.2 (63.7).
In SWE Marathon, which evaluates long-running software engineering tasks, Kimi K3 recorded a score of 42.0, slightly ahead of Opus-4.8 (40.0) and GPT-5.6 Sol (39.8).
Other coding-related benchmarks published by the company include:
- DeepSWE: 67.6
- FrontierSWE: 81.2
- Kimi Code Bench 2.0 (Internal): 72.9
Performance Across AI Agent Tasks
Kimi.ai also published benchmark comparisons covering general AI agent performance.
According to the company:
- Automation Bench: 30.8 (highest)
- BrowseComp: 91.2 (highest)
- SpreadsheetBench 2: 34.8 (highest)
- JobBench: 52.9 (second behind Fable 5)
- GDPval-AA v2 Elo: 1565.0
- AA-Briefcase Elo: 1458.0
These benchmarks evaluate AI systems on web browsing, spreadsheet operations, workplace automation, and enterprise task completion.
Visual Agent Capabilities
The benchmark report also included visual agent evaluations.
Kimi K3 achieved:
- CharXiv (RQ) with tool: 91.3
- Zerobench with tool (Pass@5): 41.0
These tests measure an AI model’s ability to interpret images and perform reasoning with external tools.
Frontend Code Arena Ranking
Separately, Arena.ai announced that Kimi K3 has reached the No. 1 position on the Frontend Code Arena leaderboard.
According to the published rankings, Kimi K3 scored 1,679 points, ahead of Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), GLM-5.2 Max (1,587), Claude Opus 4.8 Thinking (1,562), and Grok-4.5 (1,558).
The leaderboard evaluates AI models based on their ability to solve frontend software development tasks.
Kimi.ai said Kimi K3 has been developed for long-horizon agentic coding, advanced reasoning, and multimodal workflows, reflecting the growing focus among AI developers on autonomous software engineering and enterprise AI applications.
