Moonshot Kimi K2.6 ships at 1T parameters with 32B active, 256K context, and native image and video understanding — April 20 release reframes Chinese open-weight ceiling
Moonshot AI shipped Kimi K2.6 on April 20: 1 trillion total parameters with 32B active per inference pass, 256K context window, and native image and video understanding as first-class input modalities. It's the largest Chinese open-weight model released to date and reframes what "open-weight frontier" can include — multimodal, long-context, and frontier-tier scale all at once.
The 1T/32B mixture-of-experts shape is the technical headline. At those proportions, K2.6 has more raw representational capacity than any frontier-tier model with publicly-released weights (DeepSeek V4-Pro at 1.6T/49B is comparable but trades capacity for sparser activation). The 256K context window is enterprise-grade — comparable to GPT-4 Turbo's and Claude Opus 4.7's context capacity — and the native multimodal input means K2.6 is the first open-weight model that fully replaces a closed-weight frontier model on enterprise multimodal workloads.
Ant Group's inclusionAI released Ring-2.6-1T in the same window, which is the parallel Chinese open-weight push at the trillion-parameter ceiling. Combined with DeepSeek V4 and Qwen 3.7 Max, the Chinese open-weight ecosystem now has four families at frontier-tier scale, each with overlapping but distinct strengths. The market consequence is that the open-weight self-hosting customer in mid-2026 has frontier-tier choices the same customer didn't have in mid-2025. The Western closed-weight frontier (Anthropic, OpenAI, Google) now competes against a coherent open-weight competitor stack rather than against a handful of fragmented projects.
Hugging Face Blog — Best Open-Source LLM Models 2026 → · Future AGI Substack — Best LLMs in May 2026 What Actually Matters → · BentoML — Best Open-Source LLMs in 2026 →