July 27 weights shipped · 1.56TB on Hugging Face · MoonEP / FlashKDA / AgentEnv · Custom license · API $0.30–$15/M
Summary: On the evening of July 27, 2026, Moonshot AI released the full weights and technical report for Kimi K3—a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window and native vision understanding. Alongside the weights, Moonshot open-sourced three infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. The 1.56TB download hit #1 on Hugging Face's trending list within thirty minutes. Moonshot consistently calls this an "open-weight" release, never "open source"—and that distinction is the first thing worth understanding before the specs and benchmarks.
Open source vs. open weight: Headlines say "open source"; Moonshot's own materials say open weight—training data and full training code are not public
Self-hosting is not consumer-grade: 1.56TB across 896 experts; Moonshot recommends 64+ accelerator supernodes
New commercial gates: Custom license adds a $20M Model-as-a-Service revenue threshold and 100M MAU attribution—unlike the K2 Modified MIT era
Capability ≠ value pick: ~$0.95/task on the Intelligence Index—cheaper than Claude Fable 5, but pricier than GLM-5.2 (~$0.47)
Geopolitics and distillation claims: US accusations of industrial distillation landed days before weights shipped—separate technical merit from compliance risk
By the Open Source Initiative's definition, genuinely open source requires public training data, public training code, and a reproducible pipeline—not just trained weights. Kimi K3 ships weights, a technical report, and several training-adjacent infrastructure tools, but not the training data or full training code.
The license is a fully custom document with two commercial thresholds:
Model-as-a-Service revenue gate: MaaS revenue over $20 million in any trailing 12 months requires a separate agreement with Moonshot
Attribution gate: Products exceeding 100 million MAU or $20 million monthly revenue must prominently display "Kimi K3"
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion |
| Active parameters | ~104 billion |
| Architecture | Mixture-of-Experts (MoE) |
| Experts | 896 routed, 16 activated per token, plus shared experts |
| Attention | Kimi Delta Attention (KDA) + Gated MLA |
| Context window | 1,000,000 tokens |
| Multimodality | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights, MXFP8 activations |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom — Moonshot calls it "open weight," not "open source" |
| API first availability | July 16, 2026 |
| Full weights released | July 27, 2026 |
KDA replaces a scalar forgetting gate with channel-wise gating—each feature dimension gets its own decay rate. Implemented as chunkwise DPLR for linear-time recurrence; K3 interleaves KDA with Gated MLA global layers to support 1M context while keeping KV cache small.
AttnRes replaces uniform residual accumulation with input-dependent aggregation across preceding layers—~25% training efficiency gain for under 2% added cost. Pseudo-query vectors initialize to zero for stable early training.
Muon operates per attention head; MoE routes 896 experts with 16 active per token (~1.8% sparsity), using Quantile Balancing plus MoonEP for load balance.
| Technology | What it does | Key numbers |
|---|---|---|
| MoonEP | MoE communication at extreme scale | Duplicates overloaded experts; proves redundant-expert upper bound per rank |
| FlashKDA | CUTLASS KDA kernels (previously open) | 1.72×–2.22× faster prefill on H20 vs. flash-linear-attention baseline |
| AgentEnv | Firecracker microVM sandbox with KVCache.ai | Checkpoint ~133ms, resume ~49ms, up to 6.5× memory overcommit (vendor-reported) |
SWE-bench Verified (Vals AI, July 2026): Claude Opus 5 97%, GPT-5.6 Sol 96.2%, Claude Fable 5 95%, Kimi K3 93.4%.
Artificial Analysis Intelligence Index (max reasoning): Fable 5 60, GPT-5.6 Sol 59, K3 ~57 (#3 overall, #1 open-weight), GLM-5.2 51. Cost per task ~$0.95—about 60% below Fable 5 (~$2.40) but above GLM-5.2 (~$0.47). Capability ceiling of the open-weight tier, not the value pick. K3 ranks #1 on Arena.ai's Frontend Code Arena.
| Path | Best for | Barrier / cost |
|---|---|---|
Official API (kimi-k3) | Individuals, fast validation | $0.30/M cached input; $3/M miss; $15/M output |
| OpenRouter et al. | Multi-model routing | ~7 providers, mostly official pricing |
| Full weight self-host | Hyperscale callers, labs | 1.56TB + 64+ accelerator supernode |
Call https://api.moonshot.ai/v1 with model ID kimi-k3 via OpenAI-compatible SDK; test long context and cache hits
Map your workload to K3 strengths (frontend code, long docs) vs. gaps using SWE-bench and AA Index
Have legal review the Hugging Face LICENSE for MaaS and MAU gates
If self-hosting: budget 1.56TB storage and 64+ accelerators; evaluate MoonEP / FlashKDA integration
Validate Kimi Code / OpenClaw multi-model agents in a real macOS GUI session (see VNCMac below)
The release landed during WAIC 2026 and days after US accusations of industrial distillation against Anthropic's Fable model. China's Ministry of Commerce pushed back on July 28. Three days later, Alibaba unveiled Qwen3.8-Max-Preview (2.4T parameters)—pushing the open-weight race into the "3T club."
No, not by OSI standards. It's open-weight: weights, report, and some infra are public; training data and full training code are not.
Yes for most cases. MaaS over $20M trailing revenue or 100M MAU / $20M monthly revenue triggers extra terms.
64+ accelerators on a supernode. Use the API or OpenRouter for everyone else.
Trails top closed models on SWE-bench (93.4% vs. 95–97%) but is the strongest open-weight model on the market.
~3× parameters, AttnRes + Per-Head Muon, larger context and multimodal stack; new MaaS license gate.
K3 turned the largest open-weight headline into a 1.56TB download—but open weights are not open source, and most teams will still consume K3 via API. If your primary machine is Windows or Linux and you need a real macOS GUI for Kimi Code, OpenClaw routing, or iOS build baselines, buying a Mac is expensive and SSH cannot click permission dialogs. Rent a VNCMac remote Mac by the hour, validate agents over VNC, and stop when the project ends. See Mac Mini M4 plans or our July 17 K3 deep dive.
Sources: Moonshot AI blog, Hugging Face, Vals AI, Artificial Analysis, VentureBeat, Simon Willison. Verify against official data before production decisions.