guide · ai

Kimi K3

Kimi K3 (Moonshot AI, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

--- title: “Kimi K3” description: “Kimi K3 (Moonshot AI, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-kimi-k3.png imageAlt: “A retro robot representing Kimi K3” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Moonshot API quickstart — Kimi K3” url: “https://platform.kimi.ai/docs/guide/kimi-k3-quickstart” - name: “Hugging Face model card — moonshotai/Kimi-K3” url: “https://huggingface.co/moonshotai/Kimi-K3” - name: “Wikipedia — Kimi (AI)” url: “https://en.wikipedia.org/wiki/Kimi_(AI” - name: “BBC News (Francisco Velasquez, Jul 17) — Kimi K3 launch coverage” url: “https://www.bbc.com/news/articles/cy9w4q8pgp0o” - name: “NxCode (Jul 17) — Kimi K3 benchmarks & coding agent evaluation guide” url: “https://www.nxcode.io/resources/news/kimi-k3-benchmarks-coding-agent-evaluation-guide-2026” facts: - label: “Vendor” value: “Moonshot AI” - label: “Released” value: “July 27, 2026” - label: “License” value: “Open-weight (Kimi K3 License)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - kimi-k3 - ai-models - model-release draft: false --- --- title: “Kimi K3” description: “Kimi K3 (Moonshot AI, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-kimi-k3.png imageAlt: “A retro robot representing Kimi K3” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Moonshot API quickstart — Kimi K3” url: “https://platform.kimi.ai/docs/guide/kimi-k3-quickstart” - name: “Hugging Face model card — moonshotai/Kimi-K3” url: “https://huggingface.co/moonshotai/Kimi-K3” - name: “Wikipedia — Kimi (AI)” url: “https://en.wikipedia.org/wiki/Kimi_(AI” - name: “BBC News (Francisco Velasquez, Jul 17) — Kimi K3 launch coverage” url: “https://www.bbc.com/news/articles/cy9w4q8pgp0o” - name: “NxCode (Jul 17) — Kimi K3 benchmarks & coding agent evaluation guide” url: “https://www.nxcode.io/resources/news/kimi-k3-benchmarks-coding-agent-evaluation-guide-2026” facts: - label: “Vendor” value: “Moonshot AI” - label: “Released” value: “July 27, 2026” - label: “License” value: “Open-weight (Kimi K3 License)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - kimi-k3 - ai-models - model-release draft: false --- --- type: guide category: ai model: kimi-k3 vendor: Moonshot AI release_date: 2026-07-16 weights_date: 2026-07-27 status: current license: Kimi K3 License (custom — modified-MIT + revenue-share trigger at >$20M ARR) --- # Kimi K3 — Moonshot’s 2.8T open-weight MoE that ships with a real toll-booth license ## The release Moonshot AI shipped Kimi K3 in two waves: a closed/API launch on July 16, 2026 at WAIC Shanghai, then full open weights on July 27, 2026 — a 1.561 TB drop across 96 .safetensors shards on Hugging Face under a custom Kimi K3 License. It is a 2.8-trillion-parameter sparse mixture-of-experts model with 104B parameters active per token (16 of 896 experts, plus 2 shared), built on Moonshot’s new Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) hybrid stack, with native vision (MoonViT-V2, 401M) and a 1-million-token context window. Independent benchmarks (Artificial Analysis) place it at the frontier — score 57 on the Intelligence Index v4.1, ranked #4 in displayed class, on par with GPT-5.6 Sol and Claude Fable 5. The Wikipedia infobox pins it: Kimi K3 / 16 July 2026. ## Why it matters for operators K3 is the first time a self-hoster can run something frontier-class without a hyperscaler — at MXFP4 quantization, the weights fit on the upper edge of an 8×H200 or 16×B200 node. The brief is right to call it the largest open-weight model ever shipped. For ABS-style agent work the real unlock isn’t the size, it’s the workload profile: NxCode’s evaluation says K3 wins specifically on “long, valuable, tool-heavy” coding tasks, and Moonshot’s own coding harness numbers — DeepSWE 67.5, Terminal-Bench 2.1 88.3, SWE Marathon 42.0 — are the strongest frontier-class signals in the open-weight category right now. The 1M context with preserved thinking history is engineered for multi-hour, multi-step engineering sessions where context continuity matters more than response brevity. Two non-benchmark features change the calculus: (1) the automatic prompt cache — Moonshot reports >90% cache-hit rate on Kimi Code, and cache-hit input is $0.30/MTok vs $3.00 cache-miss, so the marginal cost of long agent sessions is effectively zero; (2) the revenue-share license, which is the real news. Wikipedia’s verbatim wording: requires any company with annual revenue greater than $20 million to negotiate a contract with Moonshot AI before providing Kimi K3 to external customers as a service, and any company with either monthly revenue greater than $20 million or over 100 million monthly active users to display attribution of Kimi K3 in any product that incorporates the model. Indian Express called it “open with a caveat.” AIDB called it “open with a toll-booth.” The 30% revenue-share deals with major inference providers — and the resulting ~7% OpenRouter discount ceiling — mean you cannot chase the cheapest API vendor for K3. Pick on harness quality and cache behaviour, not on price. Moonshot direct (or Kimi Code CLI) is the practical default. ## Specs that matter | Spec | Value | |---|---| | Architecture | Stable LatentMoE on Kimi Delta Attention + Attention Residuals | | Total parameters | 2.8 trillion | | Activated parameters | 104 billion (16 of 896 experts + 2 shared) | | Layers | 93 (1 dense + 92 MoE) | | Attention composition | 69 KDA + 24 Gated MLA | | Hidden dim | 7,168 (attn) / 3,584 (latent) / 3,072 per expert | | Vocab | 160K | | Context window | 1,048,576 tokens (1M) | | Modality | Text + Image (MoonViT-V2, 401M params) | | Quantization | MXFP4 weights / MXFP8 activations (QAT) | | Hardware floor | Supernode with ≥64 accelerators (Moonshot rec) | | License | Kimi K3 License (custom) | | Weights | huggingface.co/moonshotai/Kimi-K3 (96 shards) | Sources: Hugging Face model card, Moonshot API quickstart, Wikipedia Kimi (AI). ## Benchmarks The strongest independent signal is Artificial Analysis via NxCode (Jul 17, 21-min deep-dive): Intelligence Index v4.1 score 57, ranked #4 in displayed class — verdict: “near frontier, not just leading older open models.” Output speed measured at 62 tok/s (below the ~72 median), 130M output tokens across the eval, $2,690.80 total eval cost. Arena.ai ranked K3 #1 in web interface engineering in blind human preference (BBC, Jul 17). Moonshot-reported coding numbers (mostly via Kimi Code harness, max-effort): - DeepSWE: 67.5 (Kimi Code) / 67.3 (mini-SWE-agent) - ProgramBench: 77.8 raw hidden-test pass rate - Terminal-Bench 2.1: 88.3 vs 84.6 Artificial Analysis’ leading public score - FrontierSWE: 81.2 dominance - SWE Marathon: 42.0 (long-horizon, 20-task) Reasoning/knowledge (HF table, max-effort): GPQA Diamond 93.5 (vs GPT-5.6 Sol 94.1, Fable 5 92.6), AA-LCR 74.7 (highest of all listed), HLE-Full 43.5/56.0, CritPt 23.4. The non-trivial caveat from NxCode: “Most K3 coding results were published by Moonshot, and several are absent from the benchmark owners’ live boards. Mixed harnesses and maximum-effort settings reduce model-level comparability.” Cross-vendor comparison is murky. K3 was trained with preserved thinking history, so if a harness does not send earlier reasoning back correctly — or a session switches from another model mid-flight — quality can become unstable. ## Strengths - Frontier coding on long-horizon tasks. SWE Marathon 42.0 and DeepSWE 67.3-on-mini-SWE-agent are the strongest independent signals in the open-weight class. NxCode: K3 wins on “long, valuable, tool-heavy” work. - Scale + open weights + permissive-enough license for the bottom 99% — first time a self-hoster can run something frontier-class without a hyperscaler. 2.8T/104B at MXFP4 fits on the upper edge of 8×H200 or 16×B200. - 1M context + preserved thinking history. Designed for multi-hour, multi-step engineering sessions where context continuity matters more than response brevity. - Native vision + text in one model (MoonViT-V2). No separate VLM, fewer orchestration headaches. Vision input is base64 or ms://<file-id> only — no public URLs. - Automatic caching with >90% hit rate. Cache-hit at $0.30/MTok makes the marginal cost of long agent sessions effectively free. - Real frontier-class positioning. Intelligence Index 57 — on par with GPT-5.6 Sol and Claude Fable 5. ## Weaknesses - Below-median output speed (62 tok/s vs ~72 median) and premium output price ($15/MTok) → NxCode: K3 is more expensive and more verbose than comparable models in its price tier. - Harness-sensitive. Trained with preserved thinking history; cross-vendor comparison is murky because most benchmarks are Moonshot-run with their own harness. - Custom license is not actually open. The $20M-revenue threshold + 100M-MAU attribution clause is a real friction for hyperscale use, not a vibe shift. - Distillation accusations from the White House. OSTP director Michael Kratsios publicly accused Moonshot of distilling from Claude Fable (Business Insider, NPR, Jul 28). Unresolved IP overhang and possible downstream reputational risk for US-distributed services.