Is Kimi K3 Open Source? Inside Moonshot AI's 2.8-Trillion-Parameter Model

• July 16 (eve of WAIC 2026): Kimi K3 shipped on kimi.com, Kimi Work, Kimi Code, and the Kimi API — weights not yet public. Official blog title: "Kimi K3: Open Frontier Intelligence."
• July 17: Industry deep-dives on architecture; state media called it the world's largest-parameter open model.
• July 22–23: US-China "model distillation" dispute escalated. White House OSTP Director Michael Kratsios accused Moonshot of "large-scale, covert industrial distillation" of Anthropic's Fable models; Treasury Secretary Scott Bessent floated sanctions and entity-list restrictions.
• July 27 (~11 PM Beijing): Moonshot released full K3 weights, technical report, and open-sourced MoonEP and AgentEnv (FlashKDA was already public).
• July 28: China's Ministry of Commerce responded, accusing the US of "AI hegemonism" and threatening countermeasures; domestic media followed the open-weight details.

• "Open source" semantic gap: Headlines say open source; Moonshot materials never use that term — only open weight. Training data and full training code are not public, so OSI criteria are not met.
• Two license thresholds: MaaS businesses exceeding $20M in trailing-12-month revenue need a separate commercial agreement; products over 100M MAU or $20M monthly revenue must prominently display "Kimi K3" — API resellers need legal review.
• Self-hosting reality: 896 experts at 2.8T scale cannot run on consumer hardware. Moonshot recommends 64+ GPU supernodes — downloading weights does not mean you can run them locally.
• Output pricing: $15/M output beats GLM-5.2 (~$0.47/task on Intelligence Index) on capability but loses on cost-per-task — a ceiling model, not a value pick.
• Distillation dispute unresolved: Weight release did not end the White House accusation or K3 self-identifying as Claude; compliance teams still need independent assessment.
• 1.56TB download and storage: Pulling, verifying, and replicating weights is a one-time bandwidth and object-storage cost — far beyond "just change the base URL."

Kimi Delta Attention (KDA)

Standard Gated DeltaNet uses a scalar forget gate. KDA switches to per-channel gating — each feature dimension gets its own decay rate so some features persist long-term while others fade fast. A chunkwise Diagonal-Plus-Low-Rank (DPLR) recurrence runs in linear time. K3 alternates KDA layers with a small number of Gated MLA global-attention layers — the core reason 1M context is feasible without exploding KV cache cost.

Attention Residuals (AttnRes)

Classic residual connections accumulate layer outputs uniformly, diluting early-layer signal in deep stacks. AttnRes selectively aggregates all prior layer outputs based on the current input. Each layer adds only one RMSNorm and one pseudo-query vector — about 25% training efficiency gain for under 2% overhead. Pseudo-queries initialize to zero, so early training behaves like uniform averaging.

Per-Head Muon and Stable LatentMoE

The Muon optimizer is extended to independently optimize each attention head — a training-only technique invisible at API call time. The MoE layer routes 16 of 896 experts (~1.8% sparsity), with Quantile Balancing and MoonEP proving a theoretical upper bound on redundant experts to keep load balance feasible.

Moonshot did not just dump weight files — it released the engineering stack that makes training and inference stable. That is a major reason developer community reception has been strong.

Figures below prioritize independent third-party reruns; vendor self-reports are supplementary only.

Artificial Analysis Intelligence Index (max reasoning tier): Claude Fable 5 at 60, GPT-5.6 Sol at 59, Kimi K3 ~57 (#3 globally, #1 open weight), GLM-5.2 at 51, DeepSeek V4 Pro at 44. K3 costs ~$0.95/task — about 60% cheaper than Fable 5 (~$2.4) but more than GLM-5.2 (~$0.47). It is the open-weight capability ceiling, not the best value. It currently ranks #1 on Arena.ai's Frontend Code Arena.

Versus DeepSeek V4 GA: V4 wins on cost and peak-valley pricing; K3 leads on overall capability and 1M context.

Moonshot official materials never use "open source" — the consistent term is open weight. Under OSI criteria, true open source requires public training data, training code, and reproducible pipelines. K3 publishes post-training weights, a technical report, and inference infra only.

The license is no longer called Modified MIT (as in the K2 era) but a fully custom document with two commercial gates the K2 series lacked:

1. Model-as-a-Service revenue gate: If the licensee or affiliates operate MaaS and exceed $20M cumulative revenue over any trailing 12 months, a separate commercial agreement with Moonshot is required.
2. Scale branding gate: Products exceeding 100M MAU or $20M monthly revenue must prominently display "Kimi K3" in the user interface.

For most startups and SMBs, neither gate triggers. If your business plan is "compete with Moonshot's own model API," the first gate is a legal red line.

Moonshot exposes an OpenAI-compatible API (https://api.moonshot.ai/v1, model ID kimi-k3). Mooncake disaggregated inference hits 90%+ cache rates on typical coding workloads — real input cost lands closer to the $0.30 tier than the $3 headline.

Self-hosting needs 64+ GPU supernodes. For individuals and most teams, the practical path is the official API or OpenRouter (seven providers listed, pricing mostly matches official). For smaller open MoE on Apple Silicon, see our DeepSeek V4 high-memory Mac rental runbook.

Kimi K3's open-weight drop landed in the WAIC 2026 window, layered on an escalating US-China "model distillation" fight. Days before weights went public, US officials accused Moonshot of industrial-scale distillation of Anthropic models and threatened sanctions; on July 28 China's commerce ministry called it "AI hegemonism" and promised countermeasures. Releasing weights, the technical report, and three infra projects simultaneously is Moonshot's way of demonstrating engineering independence with reproducible detail. Three days later Alibaba shipped Qwen3.8-Max-Preview at 24T parameters — widely read as a direct response. Chinese frontier models have entered the "3T club."

1. 01
   Confirm license boundaries: Read the custom license; check whether MaaS revenue or MAU crosses $20M / 100M gates. Competing inference providers need legal sign-off first.
2. 02
   Pick an access path: API via https://api.moonshot.ai/v1; OpenRouter moonshotai/kimi-k3; self-host only if you have 64+ GPU supernodes.
3. 03
   Design cache-friendly prompts: Stabilize system prompts and prefix blocks so Mooncake hits the $0.30/M cache tier — target 90%+ hit rate on coding workloads.
4. 04
   Smoke-test at small scale: Run 10K-token tasks first for quality and latency; cross-check routing against OpenRouter July rankings.
5. 05
   Plan weight downloads: If fine-tuning locally, budget 1.56TB+ storage and bandwidth; verify Hugging Face file integrity before starting training pipelines.
6. 06
   Build hybrid routing and cost models: Route long-horizon coding and document analysis to K3; keep FrontierSWE-grade bug fixes on Fable 5; estimate monthly Agent CI infra on the pricing page before scaling up.

Kimi K3's full open-weight release is one of 2026's biggest events in the open-weight lane: 2.8T scale, 1M context, a complete infra stack, and #3 globally on Artificial Analysis — proof that Chinese labs can match closed frontier models on engineering innovation. But open weight is not open source, and the distillation dispute means enterprise buyers should stay clear-eyed.

Teams running K3 agents against large codebases over long sessions still need a stable, auditable local dev and CI plane. Shared minute pools, desk-side Macs, and oversubscribed VPS hosts introduce bandwidth jitter, neighbor CPU contention, and dropped long-lived connections — friction that quickly erases token savings. For production agent hosts, NUKCLOUD multi-region bare-metal Mac / cloud Mac nodes provide dedicated Apple Silicon with clear tenant boundaries. Compare specs on the pricing page and provision a trial via order.

Data as of 2026-07-28. Sources: Moonshot AI official blog, Hugging Face, Vals AI SWE-bench, Artificial Analysis Intelligence Index, VentureBeat, Scientific American. Verify against official data before production decisions.