IS KIMI K3
OPEN_
SOURCE_
OR_NOT.
Short answer: no, not by the strict definition — and Moonshot AI itself does not claim otherwise. On July 27, 2026, Moonshot released full weights, the technical report, and three infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. Pain point: "open source" and "open weight" are not the same, and the custom license adds two commercial gates most coverage ignores. Conclusion: K3 is the capability ceiling of the open-weight tier (AA Index #3 globally), not the value pick. Structure: license clarity → specs → KDA/AttnRes architecture → infra stack → benchmarks → pricing/self-hosting → geopolitical context → 5-step integration → FAQ.
1. Pain Points: Why July 27 Changes the Evaluation
1) Terminology trap: Moonshot consistently says "open weight," never "open source" — training data and full training code are not public. 2) Two license gates: Model-as-a-Service revenue above $20M over 12 months, or products exceeding 100M MAU / $20M monthly revenue, trigger additional terms. 3) Self-hosting is unrealistic: 1.56TB download, 64+ accelerator supernode recommended — API or OpenRouter is the practical path. 4) Not the cheapest option: ~$0.95/task on the Intelligence Index vs GLM-5.2 at ~$0.47. 5) Geopolitical overlay: US distillation allegations and China's July 28 counter-response add compliance and reputational risk to vendor selection.
2. Timeline: API Launch to Full Weights in 11 Days
| Date | Event |
|---|---|
| July 16 | K3 launches on kimi.com, Kimi Work, Kimi Code, and API — weights withheld |
| July 17 | Industry architecture analysis; state media calls it largest open model by parameter count |
| July 22–23 | US officials accuse Moonshot of industrial-scale distillation from Anthropic's Fable model |
| July 27, 23:00 | Full weights, technical report, MoonEP and AgentEnv open-sourced (FlashKDA already public) |
| July 28 | China's Ministry of Commerce responds; domestic media covers release details |
The 1.56TB Hugging Face release hit #1 on the trending list within thirty minutes — just 11 days after the API-only launch.
3. Kimi K3 at a Glance
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion |
| Active parameters | ~104 billion |
| Architecture | Mixture-of-Experts (MoE) |
| Experts | 896 routed, 16 activated per token + 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 (QAT from SFT stage) |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom — "open weight," not "open source" |
4. Architecture: KDA, Attention Residuals, Per-Head Muon
Kimi Delta Attention (KDA)
Standard Gated DeltaNet applies a single scalar forgetting gate. KDA replaces it with channel-wise gating — each feature dimension gets its own decay rate. Implemented as chunkwise DPLR recurrence, interleaved with Gated MLA global-attention layers. This hybrid design supports a 1M-token context while keeping KV cache size low.
Attention Residuals (AttnRes)
Uniform residual accumulation dilutes early-layer information at depth. AttnRes uses selective, input-dependent aggregation across preceding layers — one RMSNorm and one pseudo-query vector per layer, delivering ~25% higher training efficiency for under 2% added parameter cost.
Stable LatentMoE
896 experts, 16 activated per token (~1.8% sparsity). Quantile Balancing plus MoonEP addresses expert load imbalance with a proven upper bound on redundant experts per rank.
5. Three Infrastructure Releases
| Technology | Role | Key numbers |
|---|---|---|
| MoonEP | MoE communication at extreme scale | Duplicates overloaded experts temporarily; proves redundant-expert upper bound per rank |
| FlashKDA | CUTLASS-based KDA kernel | 1.72×–2.22× faster prefill on H20 vs flash-linear-attention baseline; drop-in via chunk_kda |
| AgentEnv | Firecracker microVM agent sandbox (with KVCache.ai) | Checkpoint 133ms, resume 49ms, up to 6.5× memory overcommit (vendor-reported, not independently verified) |
6. Benchmarks vs GPT-5.6, Claude, and GLM-5.2
SWE-bench Verified (Vals AI independent evaluation, July 2026)
| Model | Score | Release |
|---|---|---|
| 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 open-weight) > GLM-5.2 (51) > DeepSeek V4 Pro (44).
~$0.95/task on the Intelligence Index — ~60% cheaper than Claude Fable 5 (~$2.40) but more expensive than 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.
7. License: Two Commercial Gates
- MaaS revenue gate: Model-as-a-Service businesses generating more than $20 million in aggregate revenue over any trailing 12 months require a separate commercial agreement with Moonshot.
- Attribution gate: Products exceeding 100 million MAU or $20 million monthly revenue must prominently display "Kimi K3" in the UI.
For nearly every startup and mid-size company, neither gate is a practical concern. The first clause matters only if your business model is reselling K3 inference at scale in direct competition with Moonshot's API.
8. API Pricing and Self-Hosting
| Token type | Price per million tokens |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (incl. reasoning trace) | $15.00 |
Endpoint: https://api.moonshot.ai/v1, model ID kimi-k3, OpenAI SDK compatible. Mooncake disaggregated serving reportedly holds 90%+ cache hit rates on coding workloads — real-world input cost lands closer to $0.30 than $3.00. Self-hosting requires a 64+ accelerator supernode; OpenRouter already lists seven providers.
9. Five Steps to Integrate K3
- Pick your path: API vs third-party hosting vs self-deploy — 99% of teams should choose the first two.
- Audit the license: Check MaaS revenue and MAU thresholds against your business model.
- Wire OpenAI-compatible clients: Cursor, OpenClaw, or custom agents — change base URL and API key only.
- Exploit cache hits: Repeated code context on Mooncake architecture drives effective cost toward the $0.30 tier.
- Build a fallback chain: Route hard tasks to K3, daily traffic to GLM-5.2 or DeepSeek V4 Flash to control spend.
10. Deep Dive: WAIC 2026 and the "3T Club" Race
K3's release landed during WAIC 2026 and days after the US-China distillation dispute escalated — White House adviser Michael Kratsios accused Moonshot of industrial-scale distillation from Anthropic's Fable model; China's Ministry of Commerce pushed back on July 28, accusing Washington of "AI hegemonism." 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.
Three days later, Alibaba — a Moonshot investor — unveiled Qwen3.8-Max-Preview at 2.4 trillion parameters, widely read as a direct response. China's open-weight race has entered what some now call the "3T club" — parameter scale, architectural innovation, and infra open-sourcing moving in lockstep.
For Mac developers, self-hosting K3 is not realistic, but API access is trivial. Local MLX fallbacks with quantized GLM-5.2-class models remain viable on Apple Silicon. K3's real value: proving open-weight models can approach closed frontier capability, shifting pricing and benchmark baselines for the entire ecosystem.
11. FAQ
Is Kimi K3 open source? No, not by the OSI definition. Open-weight: trained weights, technical report, and some infra tools are public; training data and full training code are not.
Can I use it commercially? Yes, for the vast majority of use cases. Restrictions apply only at the MaaS revenue and scale thresholds described above.
How much hardware to self-host? Moonshot recommends 64+ accelerators on a supernode. Most developers should use the official API or OpenRouter.
How does K3 differ from K2? ~3× K2.5's parameter count, adds AttnRes and Per-Head Muon, expands context and multimodal capability; license adds a new MaaS revenue threshold.
12. Closing: API Access + Mac Local Fallback Architecture
A Windows or Linux cloud box can call the Kimi K3 API, but it falls short on Cursor toolchain integration, OpenClaw 24/7 agent hosting, local MLX quantized fallbacks, and graphics/multimedia workflows compared to Apple Silicon Macs. K3 self-hosting needs a 64-GPU supernode; API access needs only a base URL change. If you also need local GLM-5.2 offline backup, long-context agent offload, and stable always-on operation, a three-tier stack works best: local MLX for daily quantized open models; Kimi K3 API for hard coding and reasoning; MACGPU remote Mac nodes for OpenClaw/Hermes residency and unified-memory-heavy long-context workloads — as the 3T club race heats up, predictable compute is the best hedge.