Kimi K3

The alphaXiv-hosted Kimi K3 technical report abstract introduces a 2.8T-parameter MoE with 104B activated parameters, native vision, and a 1M-token context window, built on Kimi Delta Attention and Attention Residuals for improved information flow across sequence length and model depth. Stable LatentMoE activates 16 of Post-training emphasizes general, agentic, and coding RL across multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution. Extensive evaluations show frontier-level performance on long-horizon coding, agentic, knowledge, reasoning, and vision tasks, with overall performan

Yotta Labs characterizes Kimi K3 as the biggest open-weight release ever announced: 2.8T parameters with sparse MoE (896 experts, 16 active per token, ~104B active), 1M-token context window, native text+vision, and always-on reasoning, with weights that dropped July 27 as promised. The model ships in MXFP4 with quantiz API pricing is $3/M input ($0.30 cached) and $15/M output via Moonshot. The Kimi K3 License is free for commercial use with attribution, with extra conditions only at very large scale. Self-hosting reality: K3 does not fit on one GPU or node, requiring a multi-node cluster with 1.6 TB+ aggregate GPU memory. Independent

WhatLLM's coverage details Kimi K3 as Moonshot AI's flagship multimodal reasoning model, launched July 16, 2026, as a 2.8T-total-parameter sparse MoE activating 16 of 896 routed experts per token, with a one-million-token context window, native vision, and long-horizon tool use. K3 leads Arena's blind frontend coding r The pricing breakdown is $3/M uncached input, $0.30/M cached input, and $15/M output—substantially higher than Kimi K2.6. The article recommends K3 for complex coding, large repositories, document-heavy research, and visual production while keeping a cheaper model for routine traffic. Moonshot shipped the full checkpoi

Beam AI's model directory entry documents Kimi K3's July 16, 2026 release by Moonshot AI as a 2.8T-parameter MoE with native vision, 1M-token context, 16 of 896 experts active per token, and architectural features Kimi Delta Attention and Attention Residuals. As of the July 20, 2026 review, the full weights were schedu The piece also records that on July 17, 2026, Z.ai fell roughly 28% and another Chinese AI stock about 16% as investors repriced Chinese AI developers; US tech stocks fell with Nasdaq down 1.4% and Nvidia down 2.2%, though K3 was a catalyst rather than sole cause. Artificial Analysis estimated approximately $0.94 per c

Nathan Lambert's Interconnects analysis frames Kimi K3's July 16, 2026 launch by Moonshot AI as the strongest open-weight release since DeepSeek R1. K3 is a 2.8-trillion-parameter MoE model with weights scheduled for July 27 release, and Lambert argues the open-to-closed or American-to-Chinese performance gap has narro The piece situates K3 as a scaling-execution story rather than a reasoning-pivot like DeepSeek R1, with Moonshot advancing on data, algorithms, architecture, tools, and environments. Lambert notes the article assumes Moonshot honors its weight-release commitment and reflects on ecosystem implications under both open an

BenchLM's aggregator dashboard for Kimi K3 reports a capability score of 72.5/100 (field median 50.4), placing K3 11th of 196 ranked models as of September 22, 2026. Category results: Agentic 70.2 (rank 6 of 105, 95th percentile), Coding 63.5 (rank 13 of 135, 91st percentile), Multimodal 89.4 (rank 1 of 50, 100th perce Pricing is listed as $3 input / $15 output per million tokens, $0.30 cached, with a blended input of $9. Measured speed is 37 tok/s against a field median of 91 tok/s, with first-token latency of 58.52 seconds. Context is reported at 1.05M tokens. The dashboard tracks 48 published rows across 454 benchmarks with 8 of 7

MorphLLM's specification page covers Kimi K3 as Moonshot AI's open-weight flagship: 2.8T total parameters, 104B active per token, 1M-token context, native vision, and the Kimi Delta Attention hybrid linear attention architecture with Attention Residuals. Weights shipped July 27 on Hugging Face under the Kimi K3 License The article notes K3 uses always-on thinking tokens and flat $3 input / $15 output pricing per million tokens regardless of context length. It documents community quantization sizes and hardware floors for local deployment, plus harness bugs that drop reasoning content and a fixed-sampling rule. The page was last updat

BenchLM confirms Moonshot published Kimi K3's full 2.8-trillion-parameter weights on July 27, 2026, eleven days after launch, with the Hugging Face repository holding 96 Safetensors shards, the Kimi K3 License, and the technical report. The model is described as a Mixture-of-Experts with 2.8T total parameters and 104B The release uses quantization-aware MXFP4 weights with MXFP8 activations, with vLLM, SGLang, and TokenSpeed listed as supported inference paths. Moonshot recommends at least 64 accelerators for serving, making self-hosting impractical for most teams. API pricing is confirmed at $3.00 per million cache-miss input tokens