Moonshot AI Unveils Kimi K3 Flagship Model with 2.8 Trillion Parameters, Taking on GPT-5.6

China-based AI company Moonshot AI officially launched its most powerful flagship large language model to date, Kimi K3, on July 16. This massive model, boasting 2.8 trillion parameters, not only supports an ultra-long context window of one million tokens but also demonstrates performance that rivals—and in some cases surpasses—global top-tier closed-source models such as OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 across multiple authoritative benchmarks. The launch signals that Chinese AI large models have officially entered the first tier of global frontier intelligence.

According to the company's official technical blog and social media posts, Kimi K3 is defined as "the world's first open 3-trillion-parameter-class model," with core strengths concentrated in long-form coding, complex knowledge processing, and deep reasoning. Moonshot AI emphasized that the model achieves up to 6.3x decoding acceleration in million-token contexts through an innovative Kimi Delta attention mechanism, while attention residual technology delivers approximately 25% improvement in training efficiency at less than 2% additional computational cost.

In the most intuitive performance comparisons, Kimi K3's score curves closely overlap with or exceed those of top closed-source models across multiple tracks. According to official charts, Kimi K3 outperforms both Claude Fable 5 and GPT-5.6 Sol on general intelligence benchmarks; its performance on specialized coding tests runs neck-and-neck with these two flagship models. More notably, in the BrowseComp test measuring web search capabilities, Kimi K3 achieved extremely high scores at very low inference cost, demonstrating exceptional price-performance advantages.

Independent third-party evaluation data from Artificial Analysis further corroborates K3's strength. In its published intelligence index, Kimi K3's evaluation results surpassed Claude Opus 4.8 and GPT-5.6 Terra; on both the coding index and agent index, Kimi K3 likewise outperformed Claude Opus 4.8, proving its powerful capabilities in handling complex programming tasks and multi-step agent execution. The company also demonstrated K3's potential through a 3D game demo video, showing the model can generate interactive applications containing complex 3D models and physics engines.

Regarding the specific parameter count, sources show slight discrepancies. Moonshot AI's official technical blog explicitly states 2.8 trillion parameters, while earlier reports from Chinese tech media outlet IT Home and some tech bloggers pointed to 2.5 trillion parameters—the official figure of 2.8 trillion now stands as definitive. The context window is uniformly confirmed at one million tokens, with native multimodal processing capabilities.

On commercialization and ecosystem openness, Kimi K3 adopts a dual-track strategy. Users can already experience the model directly through the chatbot product Kimi.com and the coding agent Kimi Code. Moonshot AI simultaneously announced that the complete model weights for Kimi K3 will be fully open-sourced on July 27, a move certain to shake up the global open-source community. API pricing for Kimi K3 is also highly competitive: $3 per million input tokens, as low as $0.3 for cached input hits, and $15 per million output tokens.

The launch had been widely anticipated. Earlier on July 16, the official Kimi account posted a teaser video on social media, generating significant market attention. According to Polymarket prediction market data, the probability of K3 being released within July had reached as high as 97% to 98%, reflecting strong market expectations for Moonshot AI's technological iteration.

Kimi K3's emergence represents a critical step in Moonshot AI's technology roadmap evolution from the K2 series to the K3 series. Looking back, Moonshot AI released the MoE-architecture-based Kimi K2 base model in July 2025, with one trillion total parameters and 32 billion active parameters; in September of the same year, K2 received an update extending context length from 128K to 256K. In April 2026, the Kimi K2.6 model was released, achieving comprehensive upgrades in code generation and agent cluster capabilities. Now, K3's launch not only pushes parameter scale to a new order of magnitude but also completes the catch-up to—and partial surpassing of—global first-tier models in performance.

On corporate valuation and commercialization progress, Moonshot AI is likewise in a phase of rapid ascent. Reports indicate the company recently completed a funding round at a $20 billion valuation, with the pre-money valuation for the new round already elevated to $31.5 billion. The Kimi series' annualized recurring revenue (ARR) had surpassed the $300 million mark by mid-June 2026, demonstrating strong monetization capabilities.

With K3's open-source plan announced, the global developer community will gain access to a massive model with GPT-5.6-class performance that can be freely deployed. This will not only significantly lower the barrier to deploying cutting-edge AI capabilities but may also reshape the current industry landscape dominated by closed-source models.