TL;DR: On the night of July 27, 2026, Moonshot AI published the full Kimi K3 weight bundle (~1.56TB), a technical report, and open-sourced MoonEP, FlashKDA, and AgentEnv — the first complete release of a 3-trillion-parameter-class model. It topped Hugging Face trending within thirty minutes. Moonshot calls it open weight, not open source.
This guide is for developers and technical buyers evaluating Kimi K3 licensing, self-hosting cost, and API routing. Using Moonshot's release notes, Vals AI independent retests, and OSI licensing criteria, you will learn whether K3 is truly open source, what KDA and AttnRes buy you, how it ranks on SWE-bench Verified, what the $20M MaaS gate means, and a six-step path to production.
01 Kimi K3 Full Weight Release: 11-Day Timeline and Real Selection Pain Points
Pain points before the weights landed:
- Open source vs open weight confusion: Headlines say "open source," but Moonshot never used that term — legal teams auditing against OSI standards will misread compliance risk.
- Two new commercial license gates: K3's custom license adds a $20M Model-as-a-Service revenue threshold and a 100M MAU attribution requirement — unlike K2's Modified MIT era.
- Parameter count hides activation cost: 2.8T total parameters with only 16 of 896 experts active per token means self-hosting needs a 64-accelerator supernode, not a workstation.
- Geopolitics mixed into engineering: US distillation accusations (July 22–23), China's Ministry of Commerce response (July 28), and WAIC 2026 timing overlap — separate vendor risk from architecture decisions.
Full weights arrived just 11 days after Kimi K3 first shipped on kimi.com and the Moonshot API.
- July 16 (eve of WAIC 2026): K3 launched on kimi.com, Kimi Work, Kimi Code, and the API without public weights. Official blog: "Kimi K3: Open Frontier Intelligence."
- July 17: Industry deep dives on architecture; state media called it the world's largest open model by parameter count.
- July 22–23: US-China distillation dispute escalates. White House tech advisor Michael Kratsios accused Moonshot of industrial-scale distillation of Anthropic's Fable model; Treasury Secretary Scott Bessent floated sanctions and entity-list restrictions.
- July 27 (~23:00 Beijing): Full K3 weights, technical report, MoonEP and AgentEnv open-sourced (FlashKDA was already public).
- July 28: China's Ministry of Commerce responded, calling US actions "AI hegemonism." Three days later Alibaba shipped Qwen3.8-Max-Preview at 24T parameters — the domestic race enters the "3T club" phase.
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion (2.8T) |
| Active parameters | ~104 billion per forward pass |
| Architecture | Mixture-of-Experts (MoE) |
| Expert routing | 896 routed experts, 16 active per token (plus shared experts) |
| Attention stack | Kimi Delta Attention (KDA) + Gated Multi-head Latent Attention (Gated MLA) |
| Context window | 1 million tokens |
| Multimodal | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights + MXFP8 activations (quantization-aware from SFT onward) |
| Download size | ~1.56TB on Hugging Face |
| License | Custom open-weight license (not OSI open source) |
02 Kimi K3 Architecture: KDA, AttnRes, and Three Open Infra Projects
K3 is not a brute-force parameter dump. Moonshot redesigned three pillars that have stayed static in most frontier stacks: attention, residual connections, and optimizers.
Kimi Delta Attention (KDA): Standard Gated DeltaNet uses a scalar forget gate. KDA applies per-channel gating so each feature dimension gets its own decay rate. Chunkwise Diagonal-Plus-Low-Rank (DPLR) recursion runs in linear time. K3 alternates KDA layers with a small number of Gated MLA global-attention layers — enabling 1M-token context while keeping KV cache overhead extremely low.
Attention Residuals (AttnRes): Classic residual paths accumulate uniformly layer by layer, diluting early signals in deep stacks. AttnRes selectively aggregates prior layer outputs based on the current input. Each layer adds only one RMSNorm and one pseudo-query vector — roughly 25% training efficiency gain for under 2% overhead.
Per-Head Muon: Each attention head gets independent optimizer dynamics and adaptive convergence paths. Pure training-time engineering — invisible at API call time — but it helps K3 punch near closed-source frontier quality with fewer active parameters.
Stable LatentMoE: 896 experts, 16 active (~1.8% sparsity), Quantile Balancing for load equity, and MoonEP proving theoretical upper bounds on redundant experts per compute node.
Moonshot did not stop at weight files. Three infrastructure projects shipped alongside K3:
| Project | Role | Key capability |
|---|---|---|
| MoonEP | Ultra-scale fine-grained MoE communication library | Temporarily replicates overloaded experts to equalize tokens per node; maintains high communication efficiency under load imbalance |
| FlashKDA | CUTLASS-based KDA kernel (previously open) | 1.72–2.22× prefill speedup vs flash-linear-attention baseline on NVIDIA H20; drop-in chunk_kda backend |
| AgentEnv | Agent sandbox co-built with KVCache.ai (Firecracker microVM) | Large-scale parallel agent RL training; Moonshot reports 133ms checkpoint, 49ms restore, up to 6.5× memory oversubscription (not yet independently verified) |
Releasing weights plus MoonEP, FlashKDA, and AgentEnv signals engineering transparency — not just a checkpoint dump for Hugging Face bragging rights.
03 How Strong Is Kimi K3? SWE-bench Scores and Open-Weight License Gates
Vendor self-reported numbers are labeled separately below. Independent retests come first.
| Model | Score | Release date |
|---|---|---|
| Claude Opus 5 | 97% | 2026-07-24 |
| GPT-5.6 Sol | 96.2% | 2026-07-09 |
| Claude Fable 5 | 95% | 2026-06-09 |
| Kimi K3 | 93.4% | 2026-07-16 |
| Qwen3.7-Max | 79.4% | 2026-05-19 |
| DeepSeek-V4 | 76.2% | 2026-04-23 |
Artificial Analysis Intelligence Index (max reasoning tier): Claude Fable 5 (max + fallback) 60; GPT-5.6 Sol (max) 59; Kimi K3 (max) ~57 — #3 globally, #1 among open-weight models; GLM-5.2 (max) 51; DeepSeek V4 Pro (max) 44. K3 costs roughly $0.95 per task vs Claude Fable 5 at ~$2.4 (~60% cheaper) but more than GLM-5.2 at ~$0.47 — open-weight capability ceiling, not best price-performance. K3 also leads Arena.ai Frontend Code Arena at publication.
Open weight vs open source: OSI open source requires public training data, training code, and a reproducible pipeline. K3 publishes weights, a technical report, and inference infra — not full training data or end-to-end training code. The license is a custom document, not Modified MIT, with two commercial gates:
- $20M MaaS revenue gate: If you operate a Model-as-a-Service business and exceed $20 million cumulative revenue in any rolling 12-month period, you must sign a separate commercial agreement with Moonshot.
- 100M MAU attribution gate: If your product exceeds 100 million monthly active users, or monthly revenue exceeds $20 million, you must prominently display "Kimi K3" branding in the product UI.
Most startups and mid-market teams never hit these thresholds. If your business plan is "compete with Moonshot's own API on K3 weights," the MaaS gate is the legal red line.
| Token type | Price |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (includes reasoning) | $15.00 |
Mooncake's disaggregated inference architecture hits 90%+ cache hit rates on typical coding workloads — effective input cost often lands near the $0.30 tier. Self-hosting requires a 64-accelerator supernode minimum; most teams should use the official API or OpenRouter (seven providers listed at launch).
04 How to Use Kimi K3: API Setup and Six-Step Production Rollout
Moonshot exposes an OpenAI-compatible Chat Completions API. Model ID: kimi-k3. Most existing stacks need only a base URL and API key swap.
from openai import OpenAI
client = OpenAI(
api_key="your_moonshot_api_key",
base_url="https://api.moonshot.ai/v1"
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Review this pull request..."}]
)Primary official sources — re-open links after upstream updates:
https://huggingface.co/moonshotai
Six-step production rollout:
- Map license boundaries: Confirm your business will not exceed $20M MaaS revenue or 100M MAU in the next 12 months; if you are close, have legal review the custom license before shipping.
- Pick an access path: Direct Moonshot API at
https://api.moonshot.ai/v1or OpenRouter routemoonshotai/kimi-k3inside your existing OpenAI SDK stack. - Enable prompt caching: Configure cache keys for repo-scale context and agent loops to capture 90%+ hit rates and effective $0.30/M input pricing.
- Compare self-host vs API: 1.56TB weights need a 64-accelerator supernode; unless you already operate that cluster, API or third-party hosting is the realistic path.
- Build hybrid model routing: Route long coding and document tasks to K3; hard multi-file bug fixes can fallback to Claude Fable 5; terminal-heavy agents can keep GPT-5.6 Sol.
- Evaluate open infra: If you train MoE models or run agent RL at scale, audit MoonEP, FlashKDA, and AgentEnv against your stack — verify GitHub repos after each upstream release.
05 Citable Technical Data, FAQ, and Production Wrap-Up
Citable facts (EEAT):
- Parameter scale: 2.8T total, 896 experts with 16 active per token — first complete 3T-class open-weight release (Moonshot AI official blog, 2026-07-27).
- SWE-bench Verified: 93.4% independent retest — top open-weight score, trailing only Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 (Vals AI, July 2026).
- FlashKDA prefill: 1.72–2.22× speedup vs flash-linear-attention baseline on NVIDIA H20 (Moonshot FlashKDA GitHub — re-verify after each release).
No — strictly speaking it is open weight. Weights, a technical report, and partial inference infrastructure are public, but training data and full training code are not. That fails the OSI open-source definition.
Most commercial use is permitted under the custom license. You must negotiate with Moonshot if MaaS revenue exceeds $20M in any 12-month window, or if you exceed 100M MAU or $20M monthly revenue (attribution required).
Moonshot recommends a 64-accelerator supernode minimum for production inference on full weights. Personal hardware and most enterprise racks are not viable — API or managed hosting is the practical default.
K3 ranks ~57 on the Artificial Analysis Intelligence Index — third globally, first among open-weight models. It trails Opus 5, GPT-5.6 Sol, and Fable 5 on SWE-bench Verified but leads the open-weight field by a wide margin.
K3 is roughly 3× K2.5's parameter count, adds AttnRes and Per-Head Muon, expands context and native vision, and introduces the $20M MaaS revenue gate — tighter commercial guardrails than the Modified MIT K2 era.
Bottom line: Kimi K3's full open-weight release is a milestone for the domestic "3T club" — engineering depth (KDA, AttnRes, MoonEP) and geopolitical noise run in parallel. For enterprise buyers, read the open-weight license first, then pick API or managed hosting — downloading 1.56TB without a 64-GPU cluster is a hobby, not a production plan.
Connecting Kimi K3 to local agent workflows has hidden costs: a sleeping Mac kills long coding loops, and a 64-GPU cluster is out of reach for individual developers. For production AI coding agents, Kimi Code workflows, or OpenClaw Gateway 24/7 uptime, CALMVPS bare-metal Mac Mini rental is usually the better fit: dedicated Apple Silicon, Metal-native acceleration, flexible monthly billing, and ~120-second delivery — offload heavy agent loops to the cloud while you review locally. See pricing.