Moonshot Launches Kimi K3, 2.8T-Parameter Open Model Claiming Parity With Anthropic Opus 4.8

Moonshot has launched Kimi K3, a 2.8-trillion-parameter open-weights model that the Beijing-based company claims outperforms leading US systems from Anthropic and OpenAI in some capabilities³. The release represents the largest open-source AI model to date, significantly exceeding previous open models from Chinese competitors including DeepSeek.

Moonshot's own benchmarking positioned Kimi K3 behind only Claude Fable 5 and GPT-5.6 Sol among tested models, and ahead of Claude Opus 4.8⁴. Artificial Analysis placed Kimi K3 at 57 on its Intelligence Index and ranked it #1 on AutomationBench-AA. In Arena evaluations, Kimi K3 took the #1 spot in Frontend Code Arena with 1679 points and a 76% pairwise win rate, while ranking #9 in Text Arena with 1486 points.

The model features a 1-million-token context window, native multimodal input, and two novel architectural components: Kimi Delta Attention (KDA), which enables up to 6.3x faster decoding in million-token contexts, and Attention Residuals (AttnRes), which Moonshot said deliver approximately 25% higher training efficiency at 2% additional cost. Community observers reported the model uses LatentMoE or Stable LatentMoE with 16 activated experts out of 896, along with per-head Muon, quantile load balancing, and a new activation function called SiTU.

Moonshot said Kimi K3's design started in January 2025 and took approximately 1.5 years to reach frontier scale. The model is live on Kimi.com, Kimi Work, Kimi Code, and API, with open weights promised by July 27, 2026. Moonshot highlighted the model for long-horizon agentic coding, self-evolving workflows, and vision-in-the-loop coding and game-building workflows that iterate between code and screenshots.

Pricing was reported at $3 per 1M input tokens and $15 per 1M output tokens, with cached input discounted 90% to $0.30 per 1M tokens. Artificial Analysis estimated an average cost of $0.94 per Intelligence Index task. The latent.space headline characterized this as "Opus 4.8-class at Sonnet 5 pricing".

On the infrastructure side, vLLM said support for Kimi K3 was available on day zero, and Moonshot contributed a KDA prefix-caching implementation directly to vLLM. Moonshot's blog reportedly recommends deploying the model on supernode configurations with 64 or more accelerators for best inference efficiency. Early live serving observations put throughput at approximately 28 tokens per second via Moonshot API on OpenRouter. Artificial Analysis reported Kimi K3 used 21% fewer output tokens than K2.6 across the full Intelligence Index run.

Moonshot is also said to be raising fresh capital in a round that would value it at $31.5 billion². The company raised $2 billion in May at a $20 billion valuation.

ANALYSIS The combination of open weights, competitive benchmark placement against closed-source frontier models, and aggressive API pricing creates direct commercial pressure on Anthropic and OpenAI. ANALYSIS At the same time, the recommended deployment on supernode configurations with 64 or more accelerators suggests practical self-hosting of the full model may remain limited to well-resourced organizations.