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

Pain Points: Three Things You Must Reassess After K3's Full Release

1. Open source vs. open weight confusion: Chinese media widely calls K3 "fully open source," but Moonshot's own materials never use "open source" — they consistently say "open weight." If your legal team audits against OSI standards, you may misclassify the model: training data and full training code were not released.
2. Two new commercial license gates: K3 no longer uses the K2-era Modified MIT license. It's a fully custom document with thresholds that didn't exist before — if you run a Model-as-a-Service business exceeding $20M in trailing-12-month revenue, or your product exceeds 100M MAU, you need legal review before shipping.
3. Self-hosting gap vs. API reality: 2.8 trillion parameters, roughly 1.56TB of weights, and Moonshot's recommendation of 64+ accelerators mean "weights are public" does not mean "you can run it locally." For most teams, API or OpenRouter remains the practical path.

Is Kimi K3 Open Source, or Just Open Weight?

By the Open Source Initiative's (OSI) definition, a genuinely open-source model requires public training data, public training code, and a reproducible pipeline — not just the trained weights. Kimi K3 ships the weights, the technical report, and several training-adjacent infrastructure tools, but not the training data or the full training code. That makes it open-weight: anyone can download, run, fine-tune, and commercially deploy the model, but no one outside Moonshot can reproduce it from scratch.

The license itself has also tightened compared to the K2 family. It's no longer branded "Modified MIT" — it's a fully custom document with two commercial thresholds that didn't exist before:

1. Model-as-a-Service revenue gate. If you or your affiliates run a Model-as-a-Service business built on K3, and that business generates more than $20 million in aggregate revenue over any trailing 12 months, you must negotiate a separate commercial agreement with Moonshot AI before continuing.
2. Attribution gate. If your product or service exceeds 100 million monthly active users or $20 million in monthly revenue, you're required to prominently display "Kimi K3" in the product's user interface.

For nearly every startup, indie developer, or mid-size company, neither threshold is a practical concern — you can build and ship commercial products on K3 today without issue. The clause that actually matters is the first one, and only if your business model is reselling K3 inference at scale in direct competition with Moonshot's own API.

Kimi K3 at a Glance

Alongside the weights, Moonshot open-sourced three infrastructure technologies that powered K3's training: MoonEP, FlashKDA, and AgentEnv. The release makes K3 the largest open-weight model to ever ship in full — a 1.56TB download that hit #1 on Hugging Face's trending list within thirty minutes. Moonshot consistently calls this an "open-weight" release, never "open source."

The Architecture: 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): Forgetting, One Channel at a Time

Standard Gated DeltaNet applies a single scalar forgetting gate to an entire memory state — the whole thing decays at one rate. KDA replaces that with channel-wise gating: every feature dimension gets its own independent decay rate, so the model can hold onto some features indefinitely while aggressively forgetting others. It's implemented as a chunkwise Diagonal-Plus-Low-Rank (DPLR) formulation, which keeps the recurrence linear in time while staying hardware-efficient. K3 interleaves KDA layers with a smaller number of full global-attention (Gated MLA) layers — this hybrid design is exactly what lets it support a 1M-token context window while keeping KV cache size low.

Attention Residuals (AttnRes): Residual Connections That Choose What to Keep

In a standard deep network, residual connections accumulate every layer's output uniformly as depth increases — and the deeper the network, the more early-layer information gets diluted. AttnRes replaces uniform accumulation with selective, input-dependent aggregation across all preceding layers. The overhead is tiny: one RMSNorm and one pseudo-query vector per layer, a negligible fraction of total parameters. In testing, this delivered roughly 25% higher training efficiency for under 2% added cost. The pseudo-query vectors are initialized to zero, so the mechanism starts out equivalent to a uniform average and avoids destabilizing early training.

Per-Head Muon: Optimizing Attention Heads Independently

Muon has become a widely adopted optimizer in large-scale training. K3 extends it to operate per attention head rather than treating the whole model uniformly, giving each head a more adaptive convergence path. This is purely a training-time technique — it's invisible to anyone calling the API — but it's part of why K3 punches above its active-parameter count relative to closed frontier models.

Stable LatentMoE: Sparse by Design

K3's MoE layer routes across 896 experts and activates just 16 per token — roughly 1.8% sparsity — alongside shared experts for baseline stability. To prevent load imbalance from throttling throughput at scale, Moonshot applies Quantile Balancing and proves a mathematical upper bound on redundant experts per compute rank, guaranteeing a feasible balanced routing plan always exists.

Three Infrastructure Releases That Matter as Much as the Weights

Moonshot didn't just publish a model card — it open-sourced the infrastructure stack that made training K3 possible, which is a meaningful part of why the release landed well with developers.

Benchmarks: Kimi K3 vs. GPT-5.6, Claude, and GLM-5.2

Vendor-reported numbers vary widely across evaluation harnesses, so the figures below prioritize independent, third-party evaluations where available.

SWE-bench Verified (independent evaluation, Vals AI, as of July 2026)

Artificial Analysis Intelligence Index (neutral third-party composite score, max reasoning)

• Claude Fable 5 (max + fallback): 60
• GPT-5.6 Sol (max): 59
• Kimi K3 (max): ~57 — #3 overall, #1 among open-weight models
• GLM-5.2 (max): 51 (the previous open-weight leader)
• DeepSeek V4 Pro (max): 44

Kimi K3 is currently the strongest open-weight model on the market by raw capability, but it isn't the cheapest. It costs roughly $0.95 per task on the Intelligence Index — about 60% less than Claude Fable 5 (~$2.40/task), but still more expensive than fellow open-weight model GLM-5.2 (~$0.47/task). In short: it's the capability ceiling of the open-weight tier, not the value pick. K3 also currently ranks #1 on Arena.ai's Frontend Code Arena leaderboard.

Citable Technical Facts (EEAT)

• Scale: 2.8T total parameters, ~104B active, 1.8% sparsity (16 of 896 experts per token).
• Context and caching: 1M-token context window; Mooncake disaggregated serving reportedly holds cache hit rates above 90% on typical coding workloads.
• Independent benchmarks: Vals AI SWE-bench Verified 93.4%; Artificial Analysis Intelligence Index ~57, #1 among open-weight models, #3 overall.

Pricing and Self-Hosting: Can You Actually Run This Yourself?

Via API

Moonshot exposes an OpenAI SDK-compatible Chat Completions API at https://api.moonshot.ai/v1 under the model ID kimi-k3 — most existing integrations only need a base URL and API key change.

Because Moonshot's disaggregated Mooncake serving architecture reportedly holds cache hit rates above 90% on typical coding workloads, most real-world usage lands closer to the $0.30 rate than the $3.00 headline figure.

Self-hosting

At 2.8 trillion parameters across 896 experts, K3 is out of reach for consumer or even most prosumer hardware. Moonshot recommends deploying it on supernode configurations with 64 or more accelerators. For individual developers and most companies, the practical path is calling K3 through the official API or a router like OpenRouter, where seven providers already host it, most at pricing matching Moonshot's own.

The Bigger Picture: Kimi K3 Amid the US-China AI Dispute

Kimi K3's open-weight release didn't happen in a vacuum. It landed during the window of the World Artificial Intelligence Conference (WAIC 2026) and just days after the US-China "model distillation" dispute escalated sharply: White House science and technology adviser Michael Kratsios accused Moonshot of running a "large-scale, covert industrial distillation" campaign against Anthropic's Fable model to train K3, and US Treasury Secretary Scott Bessent floated sanctions and an Entity List designation in response. China's Ministry of Commerce publicly pushed back on July 28, accusing Washington of "AI hegemonism" and threatening countermeasures. Against that backdrop, releasing the full weights, technical report, and three infrastructure technologies reads as a deliberate signal — Moonshot is betting that showing its engineering work in full detail is the clearest way to answer questions about its methods. Three days after K3 shipped, Alibaba — one of Moonshot's investors — unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter model widely read as a direct response, pushing the open-weight race into what some are now calling the "3T club."

Five-Step Integration Runbook

FAQ

Is Kimi K3 open source?

No, not by the OSI's strict definition. It's open-weight: the trained weights, technical report, and some infrastructure tools are public, but the training data and full training code are not. Moonshot itself uses "open weight," never "open source," in its own materials.

Can I use Kimi K3 commercially for free?

Yes, for the vast majority of use cases. The only restrictions apply if you run a Model-as-a-Service business built on K3 generating over $20 million in revenue over any trailing 12 months (requires a separate agreement with Moonshot), or if your product exceeds 100 million monthly active users or $20 million in monthly revenue (requires displaying "Kimi K3" in your UI).

How much VRAM or hardware do I need to self-host Kimi K3?

Moonshot recommends supernode configurations with 64 or more accelerators. This is not realistic for consumer or most prosumer hardware — most developers should use the official API or a hosting provider like 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 independent SWE-bench Verified scores (93.4% vs. 95–97%) and ranks #3 overall on the Artificial Analysis Intelligence Index, but it's the strongest open-weight model available, ahead of GLM-5.2 and DeepSeek V4 Pro.

What's different between Kimi K3 and Kimi K2?

K3 has roughly 3x the parameters of K2.5, adds Attention Residuals and Per-Head Muon to the architecture, and significantly expands context window and multimodal capability. Its license also adds a new Model-as-a-Service revenue threshold not present in the K2 family, making commercial terms more granular.

Summary

Kimi K3's full open-weight release marks a new phase for Chinese frontier models — from "API available" to "engineering details auditable." It delivers real architectural innovation (KDA, AttnRes, Per-Head Muon, Stable LatentMoE), competitive benchmarks against closed frontier models, and three open infrastructure tools (MoonEP, FlashKDA, AgentEnv). But open weight is not OSI open source, and the custom license's two commercial gates deserve careful legal review.

Running a Kimi K3 agent entirely on a personal laptop or generic Linux VPS comes with real limits: sleep interrupts long loops, API keys mix with production codebases, and you cannot co-locate Apple toolchain work (Xcode, Fastlane, notarytool) on the same machine. For teams that need 7×24 unattended agents — running Kimi Code with K3 while also handling iOS CI or OpenClaw gateways — renting an M4 Mac cloud node from VPSMAC (native macOS, SSH + launchd daemons, same network segment as your dev tools) is typically a more stable, production-ready choice than a personal device or Linux VPS for hybrid model strategies.

Sources: Moonshot official blog · Kimi API docs · Vals AI SWE-bench Verified · Artificial Analysis Intelligence Index

Data as of July 28, 2026. Benchmarks prioritize Vals AI and Artificial Analysis independent evaluations. Model capabilities and pricing may change — verify against official docs before shipping.