Kimi K3 (SCX.ai)

The vLLM project published a detailed engineering post documenting Kimi K3 serving optimizations across multiple layers of the inference stack. The work targeted KDA recurrent state handling, LatentMoE routing, MXFP4 expert kernels, speculative decoding, and tensor/pipeline parallelism, with named changes including ada The post provides developers with reproducible deployment guidance for Kimi K3 on vLLM, including server start commands and configuration notes. It clarifies that prefix caching was disabled in the comparison runs because v0.27.1 had a known Kimi K3 prefix-caching issue that was subsequently fixed. The optimization eff

On July 27, 2026, Moonshot AI publicly released the weights of its flagship Kimi K3 model, giving developers direct access to the 2.8 trillion-parameter mixture-of-experts system rather than confining use to Moonshot's online service or API. The Beijing-based company had introduced Kimi K3 on July 16, and the open-weig According to the report, Kimi K3 has a one-million-token context window with native multimodal support across text and images, and is designed for lengthy coding, research and knowledge-work assignments. Moonshot priced K3's API at $3 per million input tokens, $15 per million output tokens and $0.30 per million cached

An Eastern Herald technology desk article reports that Moonshot AI unveiled Kimi K3 on Thursday (the publication date is July 18, 2026), describing it as an open-weight model with more than 2.8 trillion parameters and a one-million-token context window with native vision capabilities. The piece names an architecture ca The article situates K3 as a general-purpose frontier system rather than a benchmark-chasing one, emphasizing real-world deployment implications such as loading an entire codebase, a legal archive, or a book-length document into a single prompt with coherent long-context handling. It notes that chip stocks fell on Nasd

Moonshot AI released the Kimi K3 model on Thursday, July 16, 2026, according to this AI Release Tracker registry page. The listing describes K3 as an open-weight model — trained weights available to download and run — and as a 2.8-trillion-parameter system with a 1,000,000-token context window. It also notes K3 followe API pricing data on the same registry page, verified against platform.kimi.ai on August 18, 2026, lists input at $3.00 per 1M tokens, cached input at $0.30 per 1M tokens, and output at $15.00 per 1M tokens, with the note that "the provider states that Kimi K3 always reasons." This is Moonshot-direct platform pricing, s

TrendForce published an independent analysis placing Moonshot AI's Kimi K3 within the broader US-China open-source AI competition, explicitly naming the model. The article reports that Moonshot opened API access for Kimi K3 on July 16, 2026 and released full weights on July 27, and provides specific architectural detai The analysis contextualizes Kimi K3 against competing Chinese open-source releases, noting that DeepSeek launched V4 Flash (284B total / 13B active MoE parameters) on July 31 and V4 Pro (1.6T total) on August 13, while Alibaba Cloud released weights for Qwen 3.8 Max (2.4T total / 95B active) on August 14. TrendForce fr