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

Only 11 days passed between K3's API debut and the full weight release. Key milestones:

• July 16 (eve of WAIC 2026): K3 launched on kimi.com, Kimi Work, Kimi Code, and the Kimi API — weights not yet public.
• 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. White House science adviser Michael Kratsios accused Moonshot of industrial-scale distillation of Anthropic's Fable model; Treasury Secretary Scott Bessent floated sanctions. See our distillation controversy breakdown.
• July 27, ~11 PM: Full weights, technical report, MoonEP and AgentEnv open-sourced (FlashKDA was already out).
• July 28: China's Ministry of Commerce pushed back, accusing Washington of "AI hegemonism."

Core pain points for builders and enterprises:

• Terminology gap: Press calls it "open source"; Moonshot consistently says open weight — training data and full training code are not public.
• License uncertainty: No longer "Modified MIT"; new MaaS revenue and attribution gates require legal review.
• Deployment barrier: 2.8T parameters across 896 experts means 64+ accelerator supernodes — not consumer hardware.
• Benchmark divergence: Vendor numbers vary; prioritize independent retests from Vals AI and Artificial Analysis.

Moonshot consistently calls this an "open-weight" release, never "open source." That distinction is the first thing worth understanding before the specs and benchmarks.

Kimi Delta Attention (KDA): forgetting one channel at a time

Standard Gated DeltaNet applies a single scalar forgetting gate to an entire memory state. KDA replaces that with channel-wise gating: every feature dimension gets its own decay rate. Implemented as chunkwise Diagonal-Plus-Low-Rank (DPLR), keeping recurrence linear in time while staying hardware-efficient. K3 interleaves KDA with Gated MLA global-attention layers — the hybrid design behind the 1M-token window and low KV cache footprint.

Attention Residuals (AttnRes): selective layer aggregation

Standard residual connections accumulate every layer's output uniformly; deeper networks dilute early-layer signal. AttnRes replaces uniform accumulation with input-dependent aggregation across all preceding layers. Overhead: one RMSNorm and one pseudo-query vector per layer. Result: roughly 25% higher training efficiency for under 2% added cost. Pseudo-queries initialize to zero, so early training starts equivalent to a uniform average.

Per-Head Muon and Stable LatentMoE

Per-Head Muon extends the Muon optimizer to operate per attention head. The MoE layer routes across 896 experts, activating 16 per token (~1.8% sparsity), with Quantile Balancing and MoonEP proving a mathematical upper bound on redundant experts per rank.

Moonshot didn't just publish weights — it open-sourced the infrastructure stack that made training K3 possible.

For research teams trying to reproduce MoE training paths, the infra release may matter as much as the weights themselves.

Vendor-reported numbers vary across harnesses. Figures below prioritize independent third-party evaluations.

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

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 $0.30 than the $3.00 headline.

The license is the most controversial part of this release. Two commercial thresholds that didn't exist in the K2 family:

• Model-as-a-Service revenue gate. If you run a MaaS business on K3 generating more than $20 million in aggregate revenue over any trailing 12 months, you must negotiate a separate agreement with Moonshot.
• Attribution gate. If your product exceeds 100 million monthly active users or $20 million in monthly revenue, you must prominently display "Kimi K3" in the UI.

For nearly every startup or mid-size company, neither threshold is a practical concern. 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 API.

Six steps to start using Kimi K3 today:

1. Try in the browser: Visit kimi.com and sign up for instant access.
2. Get an API key: Create credentials at platform.moonshot.ai and add billing.
3. Configure an OpenAI-compatible client: Set base_url="https://api.moonshot.ai/v1", model ID kimi-k3.
4. Send a test request: Verify connectivity and latency with a short prompt.
5. OpenRouter fallback: Call moonshotai/kimi-k3 — seven providers already host it at official pricing.
6. Evaluate self-hosting: Weights are on Hugging Face (~1.56TB); Moonshot recommends 64+ accelerator supernodes. Most developers should use API or managed hosting instead.

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": "Find potential bugs in this code..."}]
)

Quotable hard data (Moonshot official, Vals AI, Artificial Analysis, July 28, 2026):

• Total parameters: 2.8 trillion — largest open-weight model ever shipped in full
• MoE sparsity: 896 experts / 16 activated = 1.8% activation rate
• Download size: ~1.56TB on Hugging Face; #1 trending within 30 minutes of release
• SWE-bench Verified: 93.4% independent retest — leads the open-weight tier
• Artificial Analysis Index: ~57 — #3 globally, #1 open-weight
• FlashKDA speedup: 1.72x–2.22x prefill on H20 vs. baseline
• API cache-hit input: $0.30/M; 90%+ hit rate on coding workloads
• Self-host threshold: 64+ accelerator supernode recommended

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.

Can I use Kimi K3 commercially for free?
Yes, for the vast majority of use cases. Restrictions apply only for large-scale MaaS revenue or products exceeding 100M MAU / $20M monthly revenue.

How much hardware do I need to self-host?
Moonshot recommends supernode configurations with 64 or more accelerators. Use the official API or OpenRouter for most teams.

How does K3 compare to GPT-5.6 and Claude?
It trails Opus 5, GPT-5.6 Sol, and Fable 5 on SWE-bench (93.4% vs. 95–97%) but ranks #3 on Artificial Analysis and #1 among open-weight models.

What's different from Kimi K2?
K3 has roughly 3x K2.5's parameters, adds AttnRes and Per-Head Muon, and expands context and multimodal capability. The license adds a new MaaS revenue threshold.

K3's 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. Three days later, Alibaba unveiled Qwen3.8-Max-Preview (2.4T parameters), pushing the open-weight race into what some call the "3T club."

API access is fast, but production teams still hit three hidden costs: long-context agents running out of memory on shared VPS hosts, multi-stage coding pipelines without a stable 24/7 runner, and 1.56TB weights requiring supercompute clusters to self-host. For Kimi Code-style agents, full-repo analysis, and MCP servers in production, JEXCLOUD multi-region bare-metal Mac nodes are the more stable path: dedicated Apple Silicon unified memory, no oversubscription jitter, launchd-managed agent gateways, 120-second provisioning. See the JEXCLOUD pricing page for nodes and rates.