Moonshot AI has open-sourced the Kimi K3 model, including weights, technical report, and training infrastructure. K3 features 2.8 trillion parameters, native vision understanding, and a 1M token context window. It achieves performance close to top closed-source models in benchmarks but at a fraction of the cost.
SynthePulse Insight · AI deep reading
Version 2 · 3 sources
Moonshot AI open-sources the 2.8-trillion-parameter model Kimi K3, which scores near frontier models on benchmarks, but independent tests reveal gaps in web and math capabilities, sparking discussions on distillation techniques.
On July 27, 2026, Moonshot AI officially open-sourced the Kimi K3 model weights and technical report, along with the high-performance attention kernel FlashKDA, MoE communication library MoonEP, and AI Agent scaling tool AgentENV. Kimi K3 is a 2.8-trillion-parameter MoE model, employing KDA hybrid linear attention (Kimi Delta Attention) and attention residuals, natively supporting visual understanding with a context window of 1 million tokens.
Moonshot AI claims the new architecture improves intelligence per unit of compute by 2.5 times. Training efficiency gains come from architectural improvements: the attention mechanism mixes KDA and Gated MLA in a 3:1 ratio, using Attention Residuals instead of traditional residual connections between layers; expert routing uses Stable LatentMoE and Quantile Balancing for load balancing. In post-training, Multi-Teacher On-Policy Distillation fuses capabilities from 9 expert models into a unified model.
On popular benchmarks, Kimi K3 scores close to Claude Fable 5 and GPT-5.6 Sol. On the Artificial Analysis Intelligence Index v4.1 leaderboard, K3 ranks fourth out of 580 models with a score of 57.1; on WebDev Arena, it ranks first with an Elo score of 1678, surpassing Claude Fable 5's 1634; on Vals AI's GDP-weighted industry benchmark suite, it ranks second with a score of 74.7%.
In terms of cost, Kimi K3 costs only $2.03 per task on BrowseComp, half the price of GPT-5.6 Sol; on GDPval-AA v2, its cost is 13% lower than GPT-5.6 Sol and less than 40% of Claude Fable 5. Moonshot AI states that open-weight models can lower the barrier to intelligence access and drive innovation.
Independent tests by the UK Cyber Research Institute found Kimi K3's web capabilities far behind frontier models, with math capabilities also lagging. This contrasts with its strong benchmark performance, suggesting the model may be over-optimized for certain evaluations.
The performance gap pattern has sparked speculation about distillation techniques. Chinese models often face such accusations, but US open-source advocates increasingly view distillation as legitimate. There is currently no direct evidence that Kimi K3 uses distillation, but the independent test shortcomings provide indirect support for this inference.
The three key technologies open-sourced by Moonshot AI each have distinct focuses: MoonEP manages expert parallel communication, balancing load through redundant expert placement; FlashKDA implements high-performance KDA attention operators, achieving 1.72 to 2.22 times prefill speedup on H20; AgentENV provides a large-scale sandbox environment based on Firecracker microVM, creating a total of 51,219,741 sandboxes during training and evaluation.
Open-sourcing this infrastructure helps lower the barrier for community reproduction and deployment, but the independent test shortcomings indicate that infrastructure open source alone is insufficient to address core model capability deficiencies.
This article is based on official information from Moonshot AI, third-party benchmark reports, and independent test results. Benchmark data comes from platforms such as Artificial Analysis and WebDev Arena; independent tests were conducted by the UK Cyber Research Institute, but specific test methods and detailed data were not provided in the sources. Distillation inference is based on performance gap patterns, not direct evidence.
Kimi K3 demonstrates competitiveness in benchmarks and cost, but independent tests reveal shortcomings in web and math capabilities, along with possible reliance on distillation, highlighting the performance and credibility challenges Chinese large models face in catching up with the frontier.
Primary report
Primary source
Also covered by 2 sources