Kimi K3 (Moonshot AI)

Moonshot AI released the Kimi K3 open weights on July 26, 2026, at roughly 7:30 PM EDT — a day ahead of the previously communicated July 27 target. Confirmed specifications from Moonshot's own platform pages include 2.8 trillion parameters and a 1,048,576-token (1M) context window. The release came after the model's in Day-0 hosted access was announced by Together AI and Modal, timed to the weights drop. A follow-up architecture deep-dive documents Kimi K3's use of LatentMoE, Kimi Delta Attention (KDA), attention residuals, and NoPE, with a verified laptop existence proof (Deltafin on M1 Max at ~16 s/token). The post separates confir

On Kimi K3 Open Day, July 27, 2026, Moonshot AI released the Kimi K3 model weights, the Kimi K3 technical report, and the supporting training infrastructure: MoonEP, FlashKDA, and AgentEnv. Moonshot describes K3 as its most capable model: a 2.8-trillion-parameter Mixture-of-Experts with native visual understanding and The technical report covers the architectural components: KDA + Gated MLA mixed at a 3:1 ratio with block-level attention residuals for efficient long-context modeling; Stable LatentMoE, which activates 16 of 896 routed experts per token using SiTU-GLU and Quantile Balancing to maintain training stability under extreme

On July 16, 2026, Moonshot AI released Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts flagship with native vision, a 1,048,576-token context window, and always-on reasoning, succeeding K2.6 (April 2026) and K2.7 Code (June 2026). The launch shipped two named variants: K3 Max for chat and single-agent workloads, a The piece contextualizes K3 as a from-scratch architectural rebuild rather than a simple K2 scale-up, noting its status as the largest open-weight model shipped at the time of release. It frames the release against competitive proprietary models and discusses benchmark interpretation with explicit attribution caveats,

Eleven days after the July 16, 2026 launch, Moonshot published Kimi K3's full 2.8-trillion-parameter weights on Hugging Face on July 27, 2026, closing the launch-day caveat about an unfulfilled open-weight promise. The official repository now holds 96 Safetensors shards under the Kimi K3 License along with the technica Because communication and expert routing become part of the serving problem at this scale, Moonshot recommends at least 64 accelerators, making the direct API the practical route for most teams rather than self-hosting. The post also documents the Kimi API pricing of $3.00 per million cache-miss input tokens, $0.30 per

An independent benchmark aggregator dated September 24, 2026, assigns Kimi K3 an overall capability score of 71.9/100, placing it 11th out of 194 ranked models and 8th out of 74 verified models. K3's strongest category is Multimodal, where it ranks 1st of 50 models at the 100th percentile with a score of 89.4 across 13 The dashboard records K3's pricing at $3 input / $15 output per million tokens, with a $0.30 cached input price and $9 blended price. Measured speed is 37 tok/s with a first-token latency of 58.52 seconds, and the context window is listed at approximately 1.05 million tokens. The Knowledge category scores 68.5 (rank 14

A third-party reference page documents Kimi K3 as a 2.8-trillion-parameter mixture-of-experts system officially launched by Moonshot AI on July 16, 2026, routing each token through 16 of 896 specialized experts. The model features a 1-million-token context window with native image and video understanding, and introduce The page records that K3 weights became publicly available on Hugging Face on July 27, 2026 under a custom Kimi K3 License. Pricing is listed at $3 per million input tokens and $15 per million output tokens, with the model served via Moonshot's API, Amazon Bedrock (available since September 18), and 15+ other providers

Chinese startup Moonshot AI unveiled Kimi K3 on July 16, 2026, marking the latest version of its Kimi model family. According to Fortune's reporting, K3 boasts 2.7 trillion parameters, described as the largest open-weight large language model available at release time (compared to DeepSeek V4's 1.6 trillion parameters) Moonshot claimed K3 performed competitively with Anthropic's Fable 5, which Fortune identifies as the most advanced AI model widely available at the time of release. The launch is framed as further shrinking the performance gap between Chinese and U.S. AI models, coming as global businesses increasingly question the co