Moonshot AI releases Kimi K3 as the largest open-weights model to date with 2.8T parameters and 1M-token context

Moonshot AI officially launched Kimi K3 as a frontier-class open-weights model, positioning it near top closed models and above prior open competitors. The model features 2.8 trillion total parameters and a 1 million-token context window, with native multimodal input capabilities and text output.

The release includes architectural innovations such as Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Moonshot claims KDA enables up to 6.3x faster decoding in million-token contexts, while AttnRes delivers approximately 25% higher training efficiency at less than 2% additional cost.

Moonshot announced that the model is live on Kimi.com, Kimi Work, Kimi Code, and API, with open weights promised by July 27, 2026. The company emphasized product positioning around long-horizon agentic coding and self-evolving workflows, as well as "vision in the loop" coding and game-building workflows that iterate between code and screenshots.

Early evaluations highlight K3's competitive performance. In Frontend Code Arena, K3 achieved a #1 ranking with 1,679 points, a 76% pairwise win rate, and topped 6 of 7 frontend domains. In Text Arena, it ranked #9 with 1,486 points, excelling in creative writing, coding, and instruction following.

Independent assessments from Artificial Analysis place K3 at 57 on the AA Intelligence Index, comparable to Opus 4.8 and GPT-5.5 but trailing Fable 5 and GPT-5.6 Sol. The model scored 1,668 Elo on GDPval v2, achieved a 53% score on AutomationBench-AA (ranking #1), and recorded 1,547 Elo on AA-Briefcase, with an average cost per task of $0.94.

Community and technical discussions around the launch have emphasized K3's architectural details, including the use of LatentMoE with 16 activated experts out of 896, a new activation function called SiTU, and per-head Muon. The model's design reportedly began in January 2025 and took approximately 1.5 years to reach frontier scale.

Pricing for K3 is reported at $3 per 1 million input tokens and $15 per 1 million output tokens, with cached input discounted to $0.30 per 1 million tokens. This positions it competitively against Sonnet 5 and other frontier models, with blended estimates suggesting a cost of $5.40 per 1 million tokens at an 80% input/20% output ratio.

Moonshot contributed a KDA prefix caching implementation directly to vLLM, with support available on day 0 of the official release. The company recommends deploying K3 on supernode configurations with 64+ accelerators for optimal inference efficiency.