Kimi K3

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window, as documented in the technical paper hosted on alphaXiv (arXiv 2607.24653) and attributed to Moonshot AI. The architecture combines Kimi Delta Attention ( Post-training emphasizes reinforcement learning across general, agentic, and coding domains at multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution. Infrastructure advances include algorithm-system co-design for KDA, perfectly balanced expert-parallel training with e

The Geopolitechs post republishes Moonshot AI's Kimi K3 Open Day announcement of 27 July 2026, confirming the release of the Kimi K3 model weights, the technical report, and three supporting infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. K3 is described as Moonshot's most capable model to date, a 2.8-tril Architectural specifics from the announcement include KDA combined with Gated MLA at a 3:1 ratio plus block-level attention residuals for efficient long-context modeling, and Stable LatentMoE activating 16 of 896 routed experts per token with SiTU-GLU and Quantile Balancing to maintain stability under extreme sparsity.

Tahir's Medium architecture breakdown of Kimi K3 emphasizes efficiency rather than raw parameter count, documenting that K3 has 896 total experts with only 16 active per token, yielding a 1.8% activation ratio. The piece compares this against Nemotron 3 Ultra at 4.3%, Mixtral 8x7B at 3.1%, and DeepSeek V3 at around 3.1 To address that bottleneck, the article describes K3's Stable Latent MoE design, which compresses token representations before routing to experts, alongside Kimi Delta Attention and block-level attention residuals for long-context modeling. The piece situates K3 alongside DeepSeek V3 in a lineage of Chinese mixture-of-

Nathan Lambert's Interconnects analysis treats Kimi K3 as the closest open model to the frontier since DeepSeek R1, describing a 2.8-trillion-parameter mixture-of-experts architecture whose weights were promised for 27 July 2026. The piece situates K3 within the open-versus-closed and US-versus-China capability gap, ar The article cites concrete benchmark placements from independent evaluators: rank 2 overall on the Vals AI index, rank 3 on Artificial Analysis's Intelligence Index behind Claude Fable and GPT-5.6 Sol Max while being cheaper, and rank 1 on Frontend Code Arena. Lambert also flags the possibility of adversarial distillat

Beam AI's model directory entry reviews Moonshot AI's Kimi K3, a 2.8-trillion-parameter mixture-of-experts model released on 16 July 2026 with native vision, a one-million-token context window, and 16 of 896 experts active per token. The page documents Kimi Delta Attention and Attention Residuals as the architectural c The review reports third-party benchmark results from Artificial Analysis: an Intelligence Index score of 57, a GDPval-AA v2 Elo of 1668 (above GPT-5.5 and Claude Opus 4.8, behind Claude Fable 5), a lead on AutomationBench-AA at 53%, and second place on AA-Briefcase, with an estimated cost of about $0.94 per completed

BenchLM's aggregator page for Kimi K3 records a 16 July 2026 release, reasoning capability, and a 1.05-million-token context window, with a composite capability score of 74.6 out of 100 ranking 7 of 232 tracked models as of 15 September 2026. API pricing is listed at $3 per million input tokens, $15 per million output Category percentile breakdowns place Kimi K3 at the 98th percentile for agentic tasks, 97th for coding, 96th for knowledge, and 100th for multimodal and grounded workflows, with strongest eligibility in Multimodal and Grounded at rank 1. The page flags that 48 source-displayable benchmark rows leave some tracked slots