Who this is for: engineering leads, AI developers, and compliance teams evaluating Moonshot AI's July 27 full-weight release. Bottom line: Kimi K3 is open weight, not open source under OSI — 2.8T total parameters, ~104B active, 896 experts with 16 per token, 1M context, 1.56TB on Hugging Face, released 11 days after the July 16 API launch. What you get: license gates ($20M MaaS revenue, 100M MAU attribution), architecture breakdown (KDA, AttnRes, Per-Head Muon, Stable LatentMoE), infra stack (MoonEP, FlashKDA, AgentEnv), independent SWE-bench and Artificial Analysis comparisons, API pricing, a six-step runbook, and FAQ. Pairs with our July 16 K3 launch review, distillation controversy analysis, and July OpenRouter rankings.
TL;DR — 30-second verdict
Full weights arrived 11 days after the API launch. The Hugging Face bundle is roughly 1.56TB and hit trending #1 within 30 minutes.
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion (2.8T) |
| Active parameters | ~104B per token |
| Architecture | Mixture-of-Experts (MoE) |
| Expert config | 896 routed experts, 16 active per token (+ shared experts) |
| Attention | Kimi Delta Attention (KDA) + Gated MLA global layers |
| Context window | 1 million tokens |
| Multimodal | Native vision (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights + MXFP8 activations (quantization-aware from SFT) |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom (open weight — not OSI open source) |
K3 reworks three components that have stayed largely unchanged across the transformer era — attention, residual connections, and optimizers. That is why it is not simply "more parameters stacked on DeepSeek."
Standard Gated DeltaNet uses a scalar forget gate — one decay rate for the entire memory state. KDA moves gating to per-channel granularity: each feature dimension gets its own decay rate, so some features persist and others fade quickly. KDA implements linear-time recurrence via chunkwise Diagonal-Plus-Low-Rank (DPLR) math, cutting KV cache overhead on long sequences. K3 alternates KDA layers with sparse global Gated MLA layers — the combination that makes 1M context practical.
Classic residual connections accumulate every layer's output uniformly; deep stacks dilute early-layer signal. AttnRes replaces uniform accumulation with input-dependent selective aggregation across prior layers. Overhead is minimal — one RMSNorm and a pseudo-query vector per layer — while Moonshot reports roughly 25% training efficiency gain at under 2% parameter cost.
K3 extends the Muon optimizer to run independently per attention head, giving each head a more adaptive convergence path. This is training-time machinery invisible at API call time, but it explains how ~104B active parameters compete with much larger closed models.
896 routed experts with 16 active per token (~1.8% sparsity), plus shared experts for baseline stability. Quantile Balancing and MoonEP jointly bound redundant expert counts per node — load imbalance does not collapse communication efficiency.
| Component | Role | Key metrics (vendor-reported unless noted) |
|---|---|---|
| MoonEP | Fine-grained MoE communication at supernode scale | Temporarily replicates overloaded experts to equalize tokens per node; proves theoretical upper bound on redundant experts |
| FlashKDA | CUTLASS-based KDA kernel (already open) | 1.72–2.22× prefill speedup on NVIDIA H20 vs flash-linear-attention baseline; drop-in backend for chunk_kda |
| AgentEnv | Agent sandbox with KVCache.ai (Firecracker microVM) | Checkpoint 133ms, resume 49ms, memory overcommit up to 6.5× — not independently verified |
Vendor numbers and independent reruns diverge. The table below prioritizes Vals AI's independent SWE-bench Verified rerun.
| Model | SWE-bench Verified | 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 (max + fallback) 60; GPT-5.6 Sol (max) 59; Kimi K3 (max) ~57 — #3 overall, #1 open-weight; GLM-5.2 (max) 51 (previous open-weight leader); DeepSeek V4 Pro (max) 44.
Net: K3 is the highest-capability open-weight option, not the best value. It currently leads Arena.ai's Frontend Code Arena leaderboard.
Moonshot has never used "open source" in official materials — only open weight. The custom license adds two commercial gates absent from the K2 series:
| Gate | Threshold | Requirement |
|---|---|---|
| MaaS revenue | >$20M cumulative over 12 months from model-as-a-service | Separate commercial agreement with Moonshot AI |
| Scale attribution | >100M MAU or >$20M monthly revenue | Prominent "Kimi K3" display in product UI |
For most startups and mid-size teams, neither gate triggers. Read the full license before building a competing inference API — context in our Opus 5 / K3 distillation week-in-review.
| Token type | Price per million tokens |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (includes reasoning) | $15.00 |
Moonshot exposes an OpenAI-compatible Chat Completions API at https://api.moonshot.ai/v1 with model ID kimi-k3. Mooncake's disaggregated inference architecture achieves 90%+ cache hit rates on typical coding workloads, so effective input cost often sits near the $0.30 tier.
Self-hosting requires a 64+ accelerator supernode. OpenRouter already lists K3 across seven providers — see our OpenRouter API integration guide for model IDs and fallback chains.
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": "system", "content": "You are a senior software engineer."},
{"role": "user", "content": "Refactor this function with type hints and docstrings."},
],
max_tokens=4096,
)
print(response.choices[0].message.content)
K3's weight release landed during WAIC 2026 and an escalating US-China distillation fight. Days before weights went public, US officials accused Moonshot of industrial-scale distillation from Anthropic's Fable model and floated sanctions. On July 28, China's Ministry of Commerce called it "AI hegemonism" and warned of countermeasures. Releasing weights, the technical report, and three infra projects reads partly as a transparency move — engineering detail as proof of an independent training path. Three days later Alibaba's Qwen3.8-Max-Preview at 2.4T entered what observers now call the "3T club." More background: Claude Opus 5 vs Kimi K3 distillation controversy.
For most teams, the 64-accelerator supernode requirement makes API access the only realistic path. But production Agent workflows — OpenClaw, Hermes Agent, Kilo Code wired to K3 — still carry three hidden costs: laptop sleep killing long sessions, network jitter breaking inference chains, and API keys scattered across dev machines with no rotation policy. None of these relate to K3's 93.4% SWE-bench score, but all determine whether that capability reaches production.
Running a multi-model Gateway on a laptop amplifies these problems as context grows toward K3's 1M-token window. Switching providers mid-session, losing cache state on reconnect, and debugging 429 storms from a sleeping machine waste more engineering time than the model saves.
For a more stable environment suited to AI Agent automation, MACCOME Mac cloud hosts are usually the better fit — real macOS, SSH handoff, and environment isolation keep K3 API routing and Agent Gateways running 24/7 on a dedicated instance. Plans and pricing: Mac mini cloud rental rates.
Sources: Moonshot AI official blog (kimi.com/blog/kimi-k3), Hugging Face (moonshotai org), GitHub (moonshotai/FlashKDA), Vals AI SWE-bench Verified leaderboard, Artificial Analysis Intelligence Index, VentureBeat, Simon Willison's blog. Verify against official data before citing in production docs.
FAQ
Is Kimi K3 actually open source?
No, not under OSI standards. Kimi K3 is open weight: trained weights, a technical report, and partial inference infrastructure are public, but training data and full training code are not. Moonshot consistently uses open weight — never open source. See our July 16 launch review for the initial API-only period.
Can Kimi K3 be used commercially?
Yes, for most commercial scenarios without extra terms. If you operate Model-as-a-Service with more than $20M annual revenue, or exceed 100M MAU / $20M monthly revenue, the custom license requires additional compliance steps.
How many GPUs does Kimi K3 self-hosting require?
Moonshot recommends a supernode with 64 or more accelerators. For most teams, use the official API or OpenRouter (seven providers). Details in our K3 launch review.
How strong is Kimi K3 compared to closed models?
Strongest open-weight model on Artificial Analysis (~57, #3 overall). SWE-bench Verified independent rerun: 93.4% vs Opus 5 at 97%. Cost is roughly double GLM-5.2 per task — capability ceiling, not best value. Distillation context: Opus 5 / K3 controversy.
What changed from Kimi K2 to K3?
K3 is roughly 3× K2.5's parameter count, adds Attention Residuals and Per-Head Muon, expands context and multimodal vision, and introduces the MaaS revenue and MAU attribution license gates absent from K2.
How do I run Kimi K3 agent stacks 24/7 in production?
Deploy OpenClaw, Hermes Agent, or Kilo Code on a dedicated Mac cloud host with K3 API or OpenRouter tiered routing — avoid laptop sleep interrupting million-token sessions. See MACCOME Mac cloud rental plans for M4/M4 Pro configs and monthly pricing.