Kimi K3 (US)

Kimi K3 is Moonshot AI's successor to the K2 line, announced mid-July 2026 with API access first and open weights released July 27, 2026 as promised. The sparse MoE transformer carries 896 experts with 16 active per token (roughly 104B active parameters) inside a 2.8-trillion-parameter model with a 1 million token cont The model was trained quantization-aware from the fine-tuning stage, so the released weights ship in MXFP4 without a separate lossy conversion step, with the full model still at 1.56 TB. Moonshot claims a 2.5x scaling-efficiency improvement over K2, while API pricing is $3 per million input tokens ($0.30 cached) and $1

Kimi K3 is Moonshot AI's flagship multimodal reasoning model, released as a hosted service on July 16, 2026. Its 2.8 trillion total parameters make it the first announced open model in the three-trillion-parameter class, and because it is a sparse Mixture of Experts, only 16 of 896 routed experts is used for each token The model targets long-horizon agent work such as navigating large codebases, operating terminals, using tools, interpreting screenshots, researching across large document sets, building spreadsheets and presentations, and iterating on visual output. Pricing is $3/M uncached input, $0.30/M cached input, and $15/M outpu

On the LLM Stats composite leaderboard, Kimi K3 lands at an LLM Stats Score of 52.5 with a blended price of $3.39 per million tokens, behind Claude Opus 5.5 at 59.7 but ahead of DeepSeek-V4.1-Flash at 51.4. Capability-tier placements include Top 10% in Animations (2 of 97), Search (2 of 90), Tool Calling (5 of 200), Co Performance holds up across conversation depth, with turn-1 scores around 15.6, turns 2-10 at 14.7, and turns 11-30 at 14.6, while turn-31+ has no data yet. The Quality Tracker is described as stable at +0.19 sigma over 175 votes in the last 7 days, and Kimi K3 ranks 2nd on AA-Briefcase with a 1548 Elo and 1st on MathV

Kimi K3 is Moonshot AI's flagship 2.8-trillion-parameter Mixture-of-Experts system combining Stable LatentMoE (activating roughly 16 of 896 experts per token) with Kimi Delta Attention and Attention Residuals. Released July 16, 2026, it runs with a 1,048,576-token context window and native multimodal input across text, API pricing is $0.30 per million input tokens on a cache hit, $3.00 per million on a cache miss, and $15.00 per million output tokens. Full weights were published on Hugging Face on July 26, 2026 (one day ahead of Moonshot's July 27 target) in the moonshotai/Kimi-K3-MXFP4 repository under the Kimi K3 License, with the

Kimi K3 is Moonshot AI's open-weight flagship, launched July 16, 2026 with a 2.8-trillion-parameter MoE (104B active), 1M-token context, native vision, and a new Kimi Delta Attention plus Attention Residuals architecture. Weights shipped July 27 under the Kimi K3 License as a native MXFP4 checkpoint, with the technical Moonshot reports 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.1, and 1686 Elo on GDPval-AA v2 against Claude Fable 5's 1747 and Claude Opus 4.8's 1593. The API is priced at $3 per million input tokens and $15 per million output, flat at any context length, while Artificial Analysis scores it 44 on Intelligence Inde

Kimi K3 is Moonshot AI's flagship open-weight large language model, launched on July 16, 2026, with full weights published on July 27, 2026 (96 safetensors shards totaling roughly 1.56 TB on Hugging Face). The model uses a Mixture-of-Experts design with 2.8 trillion total parameters and 104 billion active per token, pa Official API pricing is $3.00 per million input tokens, $0.30 on cached input, and $15.00 per million output tokens, with reasoning always on by default. On the Artificial Analysis Intelligence Index the model scores 57, ahead of every other open-weight model and behind only three closed ones, with GLM-5.2 next at 51.