Short answer: no, not by the strict definition — and Moonshot AI itself doesn't claim otherwise. If you are evaluating whether Kimi K3 belongs in your production stack, the July 27, 2026 open-weight release forces a reset: this is the largest open-weight model ever shipped in full — 2.8 trillion total parameters, ~104 billion active, a 1-million-token context window, and native vision — alongside three infrastructure technologies (MoonEP, FlashKDA, AgentEnv). This article covers the 11-day arc from API-only launch to full weights, the KDA/AttnRes/Per-Head Muon architecture story, independent benchmark comparisons, and two commercial license gates that did not exist in the K2 family. Bottom line up front: K3 is the capability ceiling of the open-weight tier, not the value pick — and Moonshot consistently calls it "open weight," never "open source."
01 Is Kimi K3 open source, or just open weight? Four selection pain points
Eleven days after the July 16 API launch review, the ~1.56TB weight download hit #1 on Hugging Face's trending list within thirty minutes. If you are still reading domestic coverage that says "fully open-sourced," expect these hidden costs:
- The open-weight vs open-source gap: By the Open Source Initiative's definition, genuinely open-source models require public training data, public training code, and a reproducible pipeline. Kimi K3 ships weights, a technical report, and several training-adjacent infrastructure tools — not the training data or full training code.
- Self-hosting is effectively out of reach: At 2.8 trillion parameters across 896 experts, K3 is not runnable on consumer or most prosumer hardware. Moonshot recommends 64+ accelerator supernodes. For most teams, the official API or OpenRouter is the realistic path.
- Two new commercial license gates: The license is no longer branded "Modified MIT." If you run a Model-as-a-Service business on K3 generating more than $20 million in aggregate revenue over any trailing 12 months, you need a separate agreement with Moonshot. Products exceeding 100 million MAU or $20 million monthly revenue must prominently display "Kimi K3" in the UI.
- Geopolitical and distillation uncertainty: On July 22–23, White House science adviser Michael Kratsios accused Moonshot of "large-scale, covert industrial distillation" against Anthropic's Fable model; China's Ministry of Commerce pushed back on July 28. See our Claude Opus 5 and Kimi K3 distillation coverage for context.
In one line: Moonshot is betting that showing engineering work in full detail answers questions about its methods — but open-weight is not OSI open source, and license plus deployment cost must enter your selection equation.
02 From API launch to full weights: timeline and spec sheet
Key timeline:
- July 16 (eve of WAIC 2026): Kimi K3 launched on kimi.com, Kimi Work, Kimi Code, and the API — weights not yet public. Official blog title: "Kimi K3: Open Frontier Intelligence."
- July 17: Industry architecture analysis; state media called it the world's largest open model by parameter count.
- July 22–23: US-China "model distillation" dispute escalated; Treasury Secretary Scott Bessent floated sanctions and Entity List designation.
- July 27 (~23:00): Full model weights, technical report, MoonEP and AgentEnv open-sourced (FlashKDA was already public).
- July 28: China's Ministry of Commerce responded publicly; three days later Alibaba unveiled Qwen3.8-Max-Preview (2.4T parameters), pushing the open-weight race into the "3T club."
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion |
| Active parameters | ~104 billion |
| Architecture | Mixture-of-Experts (MoE) |
| Experts | 896 routed experts, 16 activated per token, plus shared experts |
| Attention | Kimi Delta Attention (KDA) + Gated Multi-Head Latent Attention (MLA) |
| Context window | 1,000,000 tokens |
| Multimodality | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights, MXFP8 activations (quantization-aware training from SFT) |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom — Moonshot calls it "open weight," not "open source" |
| Token type | Price |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (including reasoning trace) | $15.00 |
Moonshot's disaggregated Mooncake serving architecture reportedly holds cache hit rates above 90% on typical coding workloads, so most real-world usage lands closer to the $0.30 input rate than the $3.00 headline figure.
03 What KDA, Attention Residuals, and Per-Head Muon actually do
Kimi K3 doesn't just scale up an existing recipe — Moonshot rebuilt three of deep learning's long-standing default components: attention, residual connections, and the optimizer.
Kimi Delta Attention (KDA): Standard Gated DeltaNet applies a single scalar forgetting gate. KDA replaces that with channel-wise gating — every feature dimension gets its own independent decay rate — implemented as a chunkwise Diagonal-Plus-Low-Rank (DPLR) formulation. K3 interleaves KDA layers with a smaller number of full global-attention (Gated MLA) layers, supporting a 1M-token context while keeping KV cache size low.
Attention Residuals (AttnRes): Standard residual connections accumulate every layer's output uniformly; deeper networks dilute early-layer information. AttnRes replaces uniform accumulation with selective, input-dependent aggregation — one RMSNorm and one pseudo-query vector per layer, delivering roughly 25% higher training efficiency for under 2% added cost.
Per-Head Muon: K3 extends the widely adopted Muon optimizer to operate per attention head rather than treating the whole model uniformly — a training-time technique invisible to API callers, but part of why K3 punches above its active-parameter count.
Stable LatentMoE: 896 experts, 16 activated per token (~1.8% sparsity), with Quantile Balancing and MoonEP proving a mathematical upper bound on redundant experts per compute rank.
| Technology | Role | Key capability |
|---|---|---|
| MoonEP | High-performance MoE communication at extreme scale | Temporarily duplicates overloaded experts so every rank gets equal token count; proves redundant-expert upper bound per rank |
| FlashKDA | CUTLASS-based KDA kernel (previously open-sourced) | 1.72×–2.22× faster prefill on NVIDIA H20 vs flash-linear-attention baseline; drop-in via chunk_kda |
| AgentEnv | Firecracker microVM sandbox (co-developed with KVCache.ai) | Massively parallel agentic RL training; Moonshot reports checkpoint latency as low as 133ms, resume 49ms, up to 6.5× memory overcommit (not yet independently reproduced) |
04 Six-step integration guide and license self-audit
- Map open-weight to your compliance boundary: Read the full custom license. Confirm whether your business triggers the Model-as-a-Service $20M trailing-12-month gate or the 100M MAU / $20M monthly revenue attribution gate. If so, contact Moonshot legal early.
- Choose API vs self-hosting: Teams below 64-accelerator supernodes should default to the official API (
https://api.moonshot.ai/v1, model IDkimi-k3, OpenAI SDK compatible) or OpenRouter (seven providers already host it). - Optimize for Mooncake cache hits: On coding workloads, cache hit rates above 90% are reported. Pin system prompts and long contexts to pull real input cost toward $0.30/M instead of $3.00/M.
- A/B test on OpenRouter: Benchmark against GLM-5.2 (~$0.47/task) and Claude Fable 5 (~$2.40/task) on your real tasks. K3 runs ~$0.95/task — open-weight capability ceiling, not the value pick. See the July 2026 OpenRouter rankings for usage trends.
- Evaluate agent-training infra needs: If you plan agentic RL on K3, investigate AgentEnv's Firecracker microVM snapshot/fork capabilities. FlashKDA can accelerate existing flash-linear-attention projects with no code changes.
- Prepare stable compute for production agents: Local Claude Code / Cursor Agent workflows plus cloud API calls need 24/7 uptime. Assess bare-metal Mac mini vs virtualized setups for Hypervisor overhead and Metal toolchain compatibility.
# OpenAI SDK-compatible integration
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": "Explain Kimi Delta Attention in one paragraph."}]
)
print(response.choices[0].message.content)05 Benchmarks, geopolitics, citeable data, and FAQ
Vendor-reported numbers vary widely across evaluation harnesses. The figures below prioritize independent, third-party evaluations where available.
| 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): Claude Fable 5 ~60, GPT-5.6 Sol ~59, Kimi K3 ~57 (#3 overall, #1 among open-weight models), GLM-5.2 ~51, DeepSeek V4 Pro ~44. K3 also currently ranks #1 on Arena.ai's Frontend Code Arena leaderboard.
- Download size: ~1.56TB on Hugging Face; #1 trending within thirty minutes of release.
- API-to-weights gap: 11 days (July 16 API launch → July 27 full weights).
- Per-task cost: ~$0.95/task for K3 vs ~$2.40 for Claude Fable 5 (~60% cheaper) but more than GLM-5.2 (~$0.47).
- FlashKDA speedup: 1.72×–2.22× faster prefill on NVIDIA H20 vs flash-linear-attention baseline.
Kimi K3's open-weight release landed during WAIC 2026 and days after the US-China distillation dispute escalated. Releasing full weights, the technical report, and three infrastructure technologies reads as a deliberate signal — Moonshot is betting that engineering transparency is the clearest answer to questions about its methods.
FAQ
Is Kimi K3 open source?
No, not by the OSI's strict definition. It's open-weight: trained weights, technical report, and some infrastructure tools are public, but training data and full training code are not. Moonshot uses "open weight," never "open source."
Can I use Kimi K3 commercially for free?
Yes, for the vast majority of use cases. Restrictions apply only if you run a Model-as-a-Service business on K3 generating over $20 million in trailing-12-month revenue, or if your product exceeds 100 million MAU or $20 million monthly revenue.
How much hardware do I need to self-host Kimi K3?
Moonshot recommends supernode configurations with 64 or more accelerators. Most developers should use the official API or OpenRouter instead.
How does Kimi K3 compare to GPT-5.6 and Claude?
It trails Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 on SWE-bench Verified (93.4% vs 95–97%) but ranks #3 on the Artificial Analysis Intelligence Index — the strongest open-weight model available.
What's different between Kimi K3 and Kimi K2?
K3 has roughly 3x K2.5's parameters, adds Attention Residuals and Per-Head Muon, and significantly expands context and multimodal capability. The license adds a Model-as-a-Service revenue threshold not present in K2.
Primary references below; verify against official pages before citing.
Moonshot AI official blog: Kimi K3
Hugging Face: moonshotai organization
For teams running Claude Code, Cursor Agent, or Kimi K3 API workflows in production, cloud APIs alone do not solve local toolchain, Metal acceleration, or 24/7 uptime — virtualized Mac setups carry Hypervisor overhead and iOS CI compatibility risk. For environments that need zero-loss native compute, stable iOS CI/CD, and 24/7 AI Agent automation, ZUKCLOUD bare-metal Mac mini cloud nodes are usually the better fit: dedicated Apple Silicon hardware, no Hypervisor tax, always-on, flexible daily/weekly/monthly billing. See the bare-metal architecture manifesto for the full argument.