Kimi K3

Moonshot AI's own Kimi K3 technical paper introduces the model as a 2.8-trillion-parameter Mixture-of-Experts system with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. The architecture combines Kimi Delta Attention and Attention Residuals with a Stable LatentMoE des The paper documents infrastructure co-design for KDA, balanced expert-parallel training with efficient memory management, and million-token agentic RL with persistent rollout and sandbox states to support deployment at 2.8T scale. Moonshot states that full Kimi K3 model weights, a technical report, and supporting stack

A Medium analysis from August 2, 2026 frames Kimi K3's July 16, 2026 release as a visible event that made a longer-running Chinese open-weight trend apparent rather than the start of it. The piece argues K3 compressed the buffer (measured in months) between the frontier and freely downloadable models, arguing this comp The article examines how to interpret adoption-share charts without mistaking traffic for total usage, weigh capability scores against the capital that produced them, and judge open-weights promises against real hosting costs. It identifies three specific upcoming releases that will confirm or break the current equilib

Simon Willison's July 17, 2026 newsletter confirms K3's launch with 2.8 trillion parameters, describing it as Moonshot AI's "most capable model to date" and the first "open 3T-class model" (rounding 2.8T up to 3T), taking the crown from DeepSeek's 1.6T v4 Pro. Self-reported benchmarks show K3 mostly beating Claude Opus On Artificial Analysis's private long-horizon knowledge work evaluation, K3 reaches an Elo of 1547, +732 points from K2.6 and behind only Claude Fable 5. Cost per task is $0.94, similar to GPT-5.6 Sol ($1.04), half the price of Opus 4.8 ($1.80), and higher than open-weights peers. K3 uses 21% fewer output tokens than K

This Substack relays Moonshot AI's "Kimi K3 Open Day" announcement from 27 July 2026, documenting that the Kimi K3 model weights, the Kimi K3 technical report, and the supporting infrastructure technologies MoonEP, FlashKDA, and AgentEnv were published on that date. It quotes Moonshot directly: K3 is "our most capable Beyond confirming that the previously promised 27 July 2026 open-weights date was met, the page adds quantitative framing that distinguishes K3 from prior Moonshot releases, particularly the claim that each unit of compute now produces roughly 2.5 times as much intelligence as before in compute-optimal terms. It also l

Trilogy AI's July 16, 2026 write-up confirms Kimi K3 is live through Moonshot's API, kimi.com, the Kimi mobile apps, Kimi Work, and Kimi Code, with OpenRouter providing immediate access for users without a Moonshot account. The model is described as a 2.8-trillion-parameter MoE with native vision, a 1-million-token con Independent rankings place K3 at 57 on Artificial Analysis's Intelligence Index (fourth of 189 models), behind Claude Fable 5 and two GPT-5.6 Sol reasoning settings, and ahead of Claude Opus 4.8, GPT-5.5 at xhigh, Claude Sonnet 5, and GLM-5.2. K3 leads Arena.ai's Code WebDev leaderboard at the time of publication. The

Nathan Lambert's Interconnects analysis positions Kimi K3 as the closest open model to the frontier since DeepSeek R1, arguing the open-to-closed and US-to-China performance gap has compressed from 6–9 months to roughly 3–5 months. Drawing on Moonshot's release materials, he reports K3 at #2 on the Vals AI index, #3 on The piece is high-signal third-party synthesis useful for understanding K3's ecosystem positioning, though it offers limited direct developer-facing API, pricing, or integration detail. Engineering teams should treat the open-to-closed gap estimates as conditional until independent post-release benchmark reproduction c

WhatLLM's July 20, 2026 guide confirms Kimi K3 launched as a hosted service on July 16, 2026 as Moonshot AI's flagship multimodal reasoning model, with the full checkpoint, a custom Kimi K3 License, deployment recipes, and a technical report subsequently published. The post reiterates specs: 2.8 trillion total paramete K3 is positioned as a frontier coder rather than a universal winner, with the strongest evidence being first place in Arena's blind frontend coding ranking at launch and competitive results across Moonshot's coding and agentic suites. Moonshot's own launch post states K3 still trails Claude Fable 5 and GPT-5.6 Sol over

A launch-timeline recap confirms Kimi K3 was officially released on July 16, 2026 by Moonshot AI, with a full day-one rollout spanning the Kimi app, kimi.com, the Kimi API, the Kimi Work desktop app, and the Kimi Code CLI. The model is described as a 2.8-trillion-parameter flagship using Kimi Delta Attention plus Atten Independent rankings landed on July 19, 2026 with Artificial Analysis scoring Kimi K3 at 57, placing it fourth of 189 models. The article lists a July 27, 2026 milestone for the open-weights release under what it calls a Modified MIT license, which would make K3 the largest open-weight model ever released at nearly 3x