Moonshot AI reportedly released the Kimi K3 open weights on July 27, 2026, after announcing the 2.8-trillion-parameter model eleven days earlier. The article says independent evaluators had begun testing the downloadable model, making this a substantive update beyond the initial hosted-API launch. The report cites Artificial Analysis placing K3 third overall, Vals AI second, and Frontend Code Arena first. It also gives Moonshot API pricing of $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens; stable prompt prefixes could therefore materially reduce

Latent Space's AINews roundup covers Moonshot AI's launch of Kimi K3 as a frontier-class open-weights model with 2.8T total parameters, a 1M-token context window, native multimodal input, and new architectural elements called Kimi Delta Attention (KDA) and Attention Residuals. The model went live on Kimi.com, Kimi Work Quoting Artificial Analysis, the post reports Kimi K3 scoring 57 on the Intelligence Index, comparable to Opus 4.8 and GPT-5.5 but behind Claude Fable 5 and GPT-5.6 Sol. Moonshot's launch is positioned around long-horizon agentic coding, self-evolving workflows, and a "vision in the loop" multimodal loop, with a Moonsh

Kimi K3 launched on July 16, 2026 as Moonshot AI’s flagship model for coding, agents, and end-to-end knowledge work, with access through Kimi.com, Kimi Work, Kimi Code, and the kimi-k3 API. The review characterizes it as suitable for controlled production trials rather than an immediate default-model migration. The supplied comparison lists K3 at $3 per million uncached input tokens, $15 per million output tokens, and $0.30 per million cached input tokens, alongside an Artificial Analysis Intelligence Index score of 57 and 4 of 189 on the cited comparison measure. It recommends testing K3 for cost-sensitive long-context agent

Digital Applied's playbook frames the July 17 hosted launch and Moonshot's promised July 27, 2026 open-weights drop as a 10-day preparation window for teams planning self-hosting. It confirms concrete specs: 2.8T total parameters with weights in MXFP4 and activations in MXFP8, 896 experts of which 16 are active per tok The checklist compares the upcoming K3 release to Kimi K2.7-Code's June Modified-MIT license as a precedent for what license clauses to watch, and uses DeepSeek V4's open-weights day as a benchmark for third-party hosting speed. It stresses that no LICENSE file or K3 deploy guide exists yet, so teams should qualify the

EmpirioLabs AI published a developer guide showing how to call Kimi K3 through an OpenAI-compatible chat completions endpoint, with thinking always on by default and surfaced in a separate `reasoning_content` field alongside `content`. The guide documents `reasoning_effort` levels (low/medium/high/max, max recommended Billing is pay-as-you-go per token with a small per-call fee only when web search runs, and Kimi K3 is positioned by EmpirioLabs as Moonshot's flagship reasoning model with a 1M-token context window suited to code assistants, autonomous agents, and multimodal analysis. The piece is explicitly AI-assisted and published

The Hacker News discussion on Moonshot's "Kimi K3: Open Frontier Intelligence" announcement reached 1999 points and 1163 comments, with Simon Willison posting a hands-on API test via OpenRouter's `moonshotai/kimi-k3` endpoint. He rendered a pelican SVG using 95 input tokens and 16,658 output tokens, of which 13,241 wer The thread links to Artificial Analysis's Kimi K3 performance and price page and surfaces community discussion of reasoning-token economics on Chinese-hosted frontier APIs. Most subsequent subthreads drift into tangential topics like LLM dungeon-master derailment rather than technical integration detail.

Moonshot AI announced Kimi K3 on July 16, 2026 as a 2.8-trillion-parameter model for coding, agents, and long-horizon work. The model was available through Moonshot’s website and API, while an open-weight release was promised for July 27, 2026. The supplied analysis cites an Artificial Analysis overall Elo of 1547 on private long-horizon knowledge work, where K3 trailed only Claude Fable 5, at a reported $0.94 per task versus $1.04 for GPT-5.6 Sol. It also reports first place on Arena.ai’s Frontend Code arena, 21% fewer output tokens than K2.6 on the Artifici

EmpirioLabs’ July 7, 2026 changelog entry introduces a Batch API that submits large asynchronous jobs at 35% below list price. Developers can upload JSONL requests to /v1/files, create a job through /v1/batches, and poll for its results; each line can target /v1/chat/completions or /v1/embeddings, with eligible models The same entry adds a Gemini-compatible endpoint. Developers can point the Python or JavaScript google-genai SDK, or another Gemini-native client, at https://api.empiriolabs.ai and authenticate with an EmpirioLabs API key to call supported chat models through the Gemini API.