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

About 22 min read · MACCOME · Last updated: July 28, 2026

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.

bolt

TL;DR — 30-second verdict

  • Not OSI open source: Moonshot uses open weight consistently — weights, a technical report, and partial infra are public; training data and full training code are not.
  • Scale ceiling: 2.8T total params, 896-expert MoE (16 active per token), 1M context, 1.56TB download — topped Hugging Face trending within 30 minutes of upload.
  • Capability positioning: Artificial Analysis Index ~57 — #3 overall, #1 open-weight — but ~$0.95/task vs GLM-5.2 at ~$0.47. Capability ceiling, not best value.
  • License gates: MaaS revenue above $20M requires a separate commercial agreement; products above 100M MAU or $20M monthly revenue must display "Kimi K3" prominently.
  • Self-host reality: Official guidance is a 64+ accelerator supernode. Most teams should use the API or OpenRouter (seven providers live).

Six Pain Points: What Teams Get Wrong After the K3 Weight Drop

  1. Calling it "open source." Headlines say open source; Moonshot's own materials say open weight. Weights, a technical report, and inference infra are public. Training data and full training code are not — that fails OSI's open source definition.
  2. Skipping the license gates. K3 ships under a custom license, not Modified MIT. Two new commercial thresholds matter if you run competing model API services or operate at hyperscale — legal review is not optional.
  3. Trusting vendor benchmarks alone. Independent reruns tell a different story. Vals AI's SWE-bench Verified rerun puts K3 at 93.4% — strong for open weights, but behind Claude Opus 5 (97%) and GPT-5.6 Sol (96.2%).
  4. Optimizing for capability, ignoring cost. K3 runs ~$0.95/task on Artificial Analysis vs GLM-5.2 at ~$0.47. It is the open-weight capability ceiling, not the price-performance winner.
  5. Underestimating self-host hardware. 2.8T parameters and 896 routed experts mean consumer hardware is out. Moonshot recommends 64+ accelerators on a supernode. API or OpenRouter is the realistic path for most teams.
  6. Running long Agent sessions on a laptop. Million-token coding agents need 24/7 Gateways. Lid-close sleep, network jitter, and scattered API keys are hidden costs unrelated to model quality but fatal to production stability.

Timeline: API Launch to Full Weights in 11 Days

  • July 16 (eve of WAIC 2026): Kimi K3 launched on kimi.com, Kimi Work, Kimi Code, and the Kimi API — online inference only, weights not yet public.
  • July 17: Industry analysis of KDA and MoE architecture began; state media reported K3 as the largest open-weight model by parameter count.
  • July 22–23: US-China distillation dispute escalated. White House tech advisor Michael Kratsios accused Moonshot of industrial-scale distillation from Anthropic's Fable model; Treasury Secretary Scott Bessent signaled possible sanctions and entity-list action.
  • July 27 (23:00 Beijing): Moonshot released full K3 weights, the technical report, and open-sourced MoonEP and AgentEnv (FlashKDA was already public).
  • July 28: China's Ministry of Commerce responded publicly, accusing the US of "AI hegemonism" and threatening countermeasures; domestic media covered release details. Alibaba shipped Qwen3.8-Max-Preview at 2.4T three days later — widely read as a direct response.

Full weights arrived 11 days after the API launch. The Hugging Face bundle is roughly 1.56TB and hit trending #1 within 30 minutes.

Kimi K3 Core Parameters

SpecValue
Total parameters2.8 trillion (2.8T)
Active parameters~104B per token
ArchitectureMixture-of-Experts (MoE)
Expert config896 routed experts, 16 active per token (+ shared experts)
AttentionKimi Delta Attention (KDA) + Gated MLA global layers
Context window1 million tokens
MultimodalNative vision (ViT-V2, 27 layers)
Weight formatMXFP4 weights + MXFP8 activations (quantization-aware from SFT)
Download size~1.56TB (Hugging Face)
LicenseCustom (open weight — not OSI open source)

Architecture: KDA, AttnRes, Per-Head Muon, and Stable LatentMoE

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."

Kimi Delta Attention (KDA)

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.

Attention Residuals (AttnRes)

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.

Per-Head Muon

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.

Stable LatentMoE

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.

Open Infra: MoonEP, FlashKDA, and AgentEnv

ComponentRoleKey metrics (vendor-reported unless noted)
MoonEPFine-grained MoE communication at supernode scaleTemporarily replicates overloaded experts to equalize tokens per node; proves theoretical upper bound on redundant experts
FlashKDACUTLASS-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
AgentEnvAgent sandbox with KVCache.ai (Firecracker microVM)Checkpoint 133ms, resume 49ms, memory overcommit up to 6.5× — not independently verified

Benchmarks: SWE-bench Verified and Artificial Analysis

Vendor numbers and independent reruns diverge. The table below prioritizes Vals AI's independent SWE-bench Verified rerun.

ModelSWE-bench VerifiedRelease
Claude Opus 597%2026-07-24
GPT-5.6 Sol96.2%2026-07-09
Claude Fable 595%2026-06-09
Kimi K393.4%2026-07-16
Qwen3.7-Max79.4%2026-05-19
DeepSeek-V476.2%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.

Open Weight vs Open Source: License Gates Explained

Moonshot has never used "open source" in official materials — only open weight. The custom license adds two commercial gates absent from the K2 series:

GateThresholdRequirement
MaaS revenue>$20M cumulative over 12 months from model-as-a-serviceSeparate commercial agreement with Moonshot AI
Scale attribution>100M MAU or >$20M monthly revenueProminent "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.

API Pricing and Self-Host Options

Token typePrice 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.

Six-Step Runbook: Deploying K3 After the Weight Release

  1. Map compliance boundaries. Read the custom license. Confirm whether your business hits the $20M MaaS revenue gate or the 100M MAU / $20M monthly revenue attribution gate. Internal use and most commercial deployments pass without extra terms.
  2. Pick an access path. Compare official API (Mooncake cache hits), OpenRouter multi-provider routing, or self-host (64+ accelerator supernode). Configure fallback chains per the OpenRouter API guide.
  3. Route by task tier. Bulk chat to cheaper Chinese flash models; complex coding and Agent work to K3 max; hardest steps to Claude Opus 5 or GPT-5.6 Sol. Follow the tiered pattern in our July OpenRouter rankings runbook.
  4. Validate on your workload. Run a SWE-bench subset or internal eval set. Compare K3 (~$0.95/task) vs GLM-5.2 (~$0.47/task) on your acceptance rate — avoid paying for capability you do not need.
  5. Deploy an Agent Gateway. Run OpenClaw, Hermes Agent, or Kilo Code on a dedicated node with K3 as the primary coding model. Configure 429/timeout failover. Keep API keys out of git.
  6. Monitor 24/7 with cost alerts. Set daily budget caps and cache-hit dashboards. Long Agent sessions belong on an always-on host — not a laptop that sleeps mid-inference.
python
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)

Three Hard Numbers for Your Technical Review

  • First fully released ~3T-class model: 2.8T total parameters, 1.56TB download — Hugging Face trending #1 within 30 minutes. No prior 3-trillion-parameter model had been released in full.
  • Independent SWE-bench Verified gap: K3 at 93.4% (Vals AI) trails Claude Opus 5 (97%) by 3.6 points but leads DeepSeek-V4 (76.2%) by 17.2 points.
  • FlashKDA prefill speedup: 1.72–2.22× on NVIDIA H20 vs flash-linear-attention baseline — long-context inference cost may drop materially once community stacks adopt the kernel.

Geopolitical Context: WAIC 2026 and the Distillation Dispute

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.

When Self-Hosting Is Not an Option: Why Agent Stacks Still Need a Dedicated Node

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.