Kimi K3 (NovitaAI)

The Marketers Index analysis (August 17, 2026) describes Kimi K3 as a 2.8 trillion parameter Mixture-of-Experts model with sparse activation: approximately 896 routed experts with only 16 engaged per input token, enabling multi-trillion-parameter capacity without prohibitive latency or cost. It highlights Kimi Delta At The article frames Kimi K3 as differentiated by API-level operational controls — specifically reasoning effort parameters and sophisticated prefix caching — positioning it as a specialized tool for high-density information processing and complex software engineering tasks rather than a pure black-box API. While the pie

The Kimi K3 architecture breakdown published on The Beat (August 17, 2026) details the model's scale-up to 2.8 trillion parameters from Kimi Linear's 48 billion, making it the largest open-weight model currently available. The article identifies several key technical innovations: a new LatentMoE component (similar to N The article positions K3 as part of a broader trend of efficient architectural changes seen in recent releases like Nemotron 3 and DeepSeek V4, framing Moonshot AI's contribution as scaling strategy combined with targeted efficiency tweaks rather than wholesale redesign. It identifies Moonshot AI as the creator of K3 a

OpenHands analyzed Kimi K3 as a sparse mixture-of-experts model with 2.8 trillion total parameters, activating 16 of 896 experts per token, and supporting text, image, and video inputs with a 1-million-token context window. The model debuted at number one on the Frontend Code Arena leaderboard, ahead of top closed models, and API pricing is listed at $0.30–$3.00 per million input tokens and $15 per million output tokens. The article noted that at the time of writing the K3 weights were announced but not yet downloadable, with Moonshot recommending deployment configurations of 64 or more accelerators and not publishing a minimum viable GPU setup. It framed Kimi K3's release alongside Thinking Machines' Inkling model as evidence that the open-weight capability gap for coding agents has narrowed significantly.

Moonshot AI released the full open weights of Kimi K3 on July 27, 2026, following its earlier announcement. The model is a native multimodal Mixture-of-Experts system with 2.8 trillion total parameters, approximately 104 billion activated per token, and a 1-million-token context window, positioned for long-horizon coding, research, and agentic work. Distribution channels include Hugging Face, GitHub, and ModelScope, accompanied by a technical report and supporting infrastructure. The release bundled architecture details, post-training strategy, long-context reinforcement-learning infrastructure, expert-parallel training systems, and inference optimizations alongside benchmark and engineering case studies. Moonshot framed the milestone as pushing open-weight scale into the 3-trillion-parameter class while disclosing substantially more than a single checkpoint, giving researchers visibility into the model's internals and deployment guidance.

An Italian-language overview published July 28, 2026 by Daniele Forciniti confirms the Kimi K3 (also called Kimi 3.0) launch on July 16, 2026, with 2.8 trillion parameters, making it the largest open-weight model ever released, surpassing DeepSeek V4 Pro at 1.6 trillion. The model is multimodal (text and image understa Kim K3 is available for free unlimited chat use on kimi.com and ranks first globally in a web-interface development benchmark ahead of Claude Fable 5. However, the API pricing makes it the most expensive Chinese model ever published. The article notes that while the model can be used free via the web interface, integra

Moonshot AI released Kimi K3 on July 16, 2026, with 2.8 trillion parameters, making it the largest open-weight AI model publicly available at launch. The model uses a Mixture-of-Experts architecture with 896 expert subnetworks, activating only 16 per request to keep inference costs manageable. It ships with a one-milli API pricing for Kimi K3 was set at $3 per million input tokens and $15 per million output tokens. On the Artificial Analysis composite leaderboard, K3 scored 732 points higher than its predecessor and finished just behind Claude Fable 5. Independent testers comparing Kimi K3 against GPT-5.6 Sol found K3 falling behind

Kimi K3 launched on July 16, 2026, as a 2.8-trillion-parameter mixture-of-experts system routing each token through 16 of 896 specialized experts, with a 1-million-token context window and native image and video understanding. Two architectural breakthroughs, Kimi Delta Attention and Attention Residuals, are credited with up to 6.3× faster long-context performance and 2.5× better scaling efficiency than Kimi K2. On the Artificial Analysis index v4.3.2, K3 scores 44 and ranks third among large open-weight models. Weights became publicly available on Hugging Face on July 27, 2026 under a custom Kimi K3 License, with the model served via Moonshot's API, Amazon Bedrock since September 18, and 15+ other providers at $3/$15 per million tokens. The model ranks first on four of eight real-world agentic benchmarks, including Program Bench, SWE Marathon, SpreadsheetBench 2, and Automation Bench, plus BrowseComp.

Moonshot AI officially opened weight downloads for Kimi K3 on July 27, 2026, making it the first open-weights release at the 2.8-trillion-parameter scale with a 1-million-token context window. The architecture replaces standard full attention with the proprietary Kimi Delta Attention hybrid linear attention mechanism alongside Attention Residuals, reducing quadratic attention complexity to linear scaling at this parameter count. Benchmark evaluations report 91% on repository bug fixing and 88% on multi-step tool invocation, highlighting performance in long-horizon reasoning and agent workflows. Moonshot simultaneously released three open-source infrastructure components: MoonEP for expert-parallel scheduling, FlashKDA for CUDA-accelerated KDA operators, and AgentEnv for long-horizon agent evaluation.