Kimi K3 is Moonshot AI's flagship: a ~2.8-trillion-parameter Mixture-of-Experts model, the first openly-committed model to reach the 3-trillion class, with a 1M-token context, native vision, and always-on reasoning. Two things make it worth a careful read rather than the hype. First, it launched hosted-only on 16 July 2026 — and the open weights shipped on 27 July, exactly as promised: 96 safetensors on moonshotai/Kimi-K3 under the permissive Kimi K3 License. K3 is now open-weight in practice, not just in commitment. Second, it is priced as a quality play, not a budget one ($3 / $15 per 1M tokens) — the deliberate break from the cheap-open-weight story its rivals tell. Where GLM competes on price, K3 went up-market and charges for it.
Moonshot AI is a Beijing lab backed by Alibaba, Tencent, and Meituan — it has reportedly raised around $3.77B across four rounds and was in talks at a ~$30B valuation (per Bloomberg, June 2026; attribute, not settled). Its consumer Kimi app is one of China's most-used AI products, with ARR reportedly passing $200M in April 2026. The model lineage worth anchoring: K2 (Jul 2025, open-weight, 1T MoE) → K2.5 (Jan 2026, multimodal) → K2.6 (Apr 2026) → K2.7-Code (Jun 2026) → K3 (16 Jul 2026).
Judge K3 the way this sub-tree judges every model: on tool-use reliability and agent-framework fit, not chat vibes. Its position in the Models band is specific — the frontier-adjacent open-weight flagship that went up-market. Where GLM-5.2 competes on price (roughly one-sixth of a Western frontier model), K3 competes on capability and prices itself alongside the Western frontier deliberately. That contrast — price-open-weight vs capability-open-weight — is the reason to read the two leaves as a pair.
K3 is the tenth model story in this band and the fourth open-weight option, alongside GLM, Llama, and Mistral / Qwen. It is the one you reach for on the hard tail — long-horizon agentic coding, large-repo navigation, native-vision-plus-1M-context work — and can absorb premium per-task cost. It is not the cheap end of a barbell; that's GLM-5.2 or K2.7-Code. The catch, throughout: as of writing you cannot download it.
K3 is text-out multimodal-in, tuned for long-horizon agentic work with a single "max" reasoning effort that is always on. The architecture below is Moonshot-disclosed; the caveats are where the honest reading lives.
Stable LatentMoE: only 16 of 896 routed experts fire per token — extreme sparsity, roughly 1.8% of parameters active, sparser than K2's 32-of-384. Quantile Balancing handles routing; a latent-space shared expert cuts deployment memory. (Exact active-param count awaits the model card.)
A hybrid linear-attention scheme. Note: the widely-quoted "~6.3× faster decoding / 75% lower KV-cache" figures originate from Moonshot's earlier Kimi Linear research paper (a 48B/3B model), not from K3 measurements — don't attach them to K3.
A drop-in replacement for standard residual connections, which Moonshot claims gives consistent scaling gains. Both KDA and AttnRes were previously published as open research.
A 1M-token window and native multimodal input (image in, text out) — long enough to hold a whole repository plus visual context in one session.
Moonshot claims a ~2.5× scaling-efficiency improvement over Kimi K2 (vendor-claimed). MXFP4 weights / MXFP8 activations, with quantization-aware training from the SFT stage — not post-training quant.
The API speaks the OpenAI format — model kimi-k3 at api.moonshot.ai/v1. Note the difference from GLM, which is Anthropic-Messages-compatible: your retarget path depends on which SDK your harness already speaks.
The drop landed on 27 July 2026, exactly as promised: moonshotai/Kimi-K3 on Hugging Face, 96 safetensors (~1.4TB at bf16), config, and a LICENSE. The shipped config.json ratifies the architecture the launch claimed: num_experts: 896, num_experts_per_token: 16, plus 2 shared experts — 16-of-896 is now a checkable file, not a vendor slide. The licence is the Kimi K3 License: permissive for use, modification, self-hosting, fine-tuning, and commercial products, with one gate — a business operating Model-as-a-Service (reselling model access) whose aggregate revenue exceeds $20M over any 12 months must sign a separate agreement with Moonshot first. End-user products with the model embedded in features are explicitly excluded from that definition, so for internal or embedded enterprise use the licence reads clean.
Even once the weights ship, this is not a rack-in-the-office model. Moonshot recommends a supernode of ≥64 accelerators to serve K3. For scale reference, the 1T K2.7-Code needs ~577GB VRAM at INT4; a 2.8T model needs considerably more. This is supernode-class inference — for most teams the honest answer is the hosted market or vLLM on serious infrastructure, not Ollama on a workstation.
Moonshot is Beijing-based; call K3's hosted API and your prompts route through servers governed by China's National Intelligence Law — a real data-path consideration for regulated or sensitive workloads. As with GLM, the risk is in the serving path, not the weights: self-hosting the open weights (once they exist) or serving through a trusted non-Moonshot host removes the exposure. Hold the same honest tension: the export-control geopolitics that makes the hosted API a governance risk is also what makes holding open weights you control — from any lab — a continuity hedge. For K3 that hedge became real on 27 July — the weights are downloadable and the licence permits self-hosting.
Read the rows honestly. Every coding benchmark in circulation is vendor-reported, max-effort, on Moonshot's own launch chart (which mixes harnesses — Kimi Code, Claude Code, Codex, mini-SWE-agent). The one independent anchor is the Artificial Analysis index. Both are below.
| Benchmark | Kimi K3 | Comparison / note | Source |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 57 | 4th of 189; behind Claude Fable 5 and two GPT-5.6 Sol settings; ahead of Claude Opus 4.8, GPT-5.5 (xhigh), Sonnet 5, GLM-5.2 | Independent (AA), Jul 2026 |
| Terminal-Bench 2.1 | 88.3 | 2nd globally (GPT-5.6 Sol 88.8); first open-weight model in the Terminal-Bench top 2 | Vendor-reported |
| FrontierSWE | 81.2 | — | Vendor-reported |
| ProgramBench | 77.8 | raw pass | Vendor-reported |
| DeepSWE | 67.5 | — | Vendor-reported |
| SWE Marathon | 42.0 | leads the field per Moonshot | Vendor-reported |
By Moonshot's own disclosure, K3's overall performance trails Claude Fable 5 and GPT-5.6 Sol, while it "substantially outperformed" Claude Opus 4.8, GPT-5.5, and Sol on selected coding/agent rows. Independent Artificial Analysis corroborates the near-frontier placement (57, 4th of 189) but also flags high output-token use, slower-than-median generation, and premium pricing. The load-bearing claim is narrow and real: an open-weight model — once shipped — is now genuinely in the top-5 frontier conversation, and it's the first Chinese lab to price like it belongs there. Do not upgrade the vendor rows into flat facts; independent SWE-bench Verified / Terminal-Bench reruns on a comparable harness don't exist yet.
The API runs $3.00 / 1M input (cache miss), $0.30 / 1M input (cache hit), $15.00 / 1M output — the highest of any Chinese lab, deliberately aligned to Western pricing. Moonshot reports >90% cache-hit rates on coding workloads (vendor-reported), which materially lowers effective input cost; consumer access is bundled in Kimi plans (~$19–$199/month). The GLM contrast is the point: this is not the "one-sixth the cost" story — it's roughly half the per-task cost of Claude Opus 4.8 by Moonshot's framing, but the always-on "max" reasoning inflates output tokens. Independent testers logged ~13,241 reasoning tokens for a single SVG generation (~$0.25/query): the list price hides the real per-task bill. Capability and operating efficiency point in opposite directions.
The economic frame is the same one the GLM leaf lays out — audit your task distribution, count the harness-rewrite cost, route with a barbell, keep a sovereign-AI backstop. Three K3-specific adjustments change the conclusion.
Unlike GLM, K3 is not the cheap end of a barbell. If you're routing center-of-distribution volume to save money, K3 is the wrong open-weight pick — reach for GLM-5.2 or K2.7-Code (roughly a quarter of the output cost). K3 earns its slot on the hard tail: long-horizon agentic coding, large-repo navigation that was pushing K2.7-Code's 256K limit, native-vision-plus-1M-context work. Buy it for capability on the difficult minority, not for price on the bulk.
Moonshot named three, and honesty is a feature: (a) thinking-history sensitivity — harnesses that truncate or modify the chain-of-thought degrade quality badly, so preserve the full trace; (b) excessive proactiveness — in ambiguous cases K3 acts rather than asks, which needs guardrails; (c) heavy reasoning-token consumption — the always-on "max" effort is the cost driver. These are real harness-design requirements before you get the benchmark behaviour.
The 27 July drop landed and the LICENSE is confirmed: the Kimi K3 License, not the K2.7 attribution formula. It is permissive for self-hosting, fine-tuning, derivatives, and embedded commercial use; the one gate is for Model-as-a-Service operators — businesses reselling K3 inference or fine-tuning — with aggregate revenue over $20M in any 12 months, who need a separate agreement with Moonshot. That is stricter than K2.7's attribution-only clause for resellers, and irrelevant for a bank self-hosting internally. The continuity hedge is now a capability, not a promise.
The tidy POPIA answer — "the data never leaves your rack" — is now real for K3: the weights shipped and the licence permits self-hosting. What hasn't changed is the hardware: ≥64-accelerator supernode territory and a ~1.4TB download, not an on-prem box most SA teams have. Watch for community quantizations (4-bit and below), which are the realistic path to bringing the entry cost down — on the pattern every large open-weight drop has followed.
For POPIA-sensitive prompts, routing to Moonshot's hosted API is a cross-border transfer through servers under China's National Intelligence Law — a data path to treat with eyes open, and to run the section 72 test against (see the Data privacy & POPIA leaf). It is not a clean residency answer.
Evaluate kimi-k3 via API for the hard-tail coding where its capability earns the premium; keep K2.7-Code or GLM-5.2 for the cheap, high-volume work; and price the self-host case on real numbers now that the repo and LICENSE exist — the open question is hardware economics, not availability. Own the context layer regardless — that's the part no model choice rescues.
moonshotai/kimi-k3 — and the routing seam for a barbell.Several statements here convert from "promised" to "measured" when the weights drop — watch the Hugging Face org for the real repo, the LICENSE, and the technical report.
kimi-k3, api.moonshot.ai/v1
platform.moonshot.ai
Weights · shipped 27 Jul
moonshotai/Kimi-K3 (weights, config, Kimi K3 License)
huggingface.co/moonshotai/Kimi-K3
Reference · architecture
OpenLM.ai — Kimi K3 architecture summary
openlm.ai/kimi-k3
Product
Kimi — the consumer app
kimi.com
As-of 29 July 2026: weights shipped 27 July, on time, to moonshotai/Kimi-K3 — Kimi K3 License confirmed, and the config ratifies 16-of-896 experts. Independent placement (Artificial Analysis index 57 — now the top open-weight result) is corroborated; the coding-benchmark rows remain vendor-reported pending the first independent SWE-bench / Terminal-Bench reruns. One independent evaluation has reported a materially higher hallucination rate than Moonshot's charts show — methodology unverified; weigh accordingly.