Is Kimi K3 open source? Inside Moonshot AI's 2.8-trillion-parameter model

Why "open weight" is not the same as open source

The friction is not whether you can download the checkpoint. It is a stack of decisions that do not show up in a Hugging Face trending badge:

• Reproducibility gap: Open Source Initiative standards require public training data, training code, and a reproducible pipeline — not just finished weights. K3 ships weights, a technical report, and inference-adjacent infra tools, but not the full training stack.
• License gates: The custom license adds two commercial thresholds absent from the K2 family — a $20M Model-as-a-Service revenue cap and a 100M MAU attribution requirement — that matter if you plan to resell inference at scale.
• Hardware reality: A 2.8T sparse MoE in MXFP4 format needs a 64+ accelerator supernode to self-host. For most teams the API or a router remains the only practical path for months.
• Geopolitical noise: The July 27 weight drop landed days after White House distillation accusations and a July 28 Commerce Ministry response — see our Kimi K3 distillation breakdown for what is verified vs disputed.
• Evaluation lag: Vendor benchmarks and third-party scores diverge. Independent SWE-bench puts K3 at 93.4%; the Artificial Analysis index ranks it ~57 overall — strong, but not a universal sweep.

Open weight vs closed frontier — license, benchmarks, and cost

Moonshot released API access on July 16 and full weights eleven days later on July 27. The table below compares what you actually get against closed alternatives and the prior Kimi K3 launch guide we published before weights landed.

K3 is the capability ceiling of the open-weight tier on independent composite scores, but not the value pick. At roughly $0.95 per task on the Artificial Analysis index it undercuts Claude Fable 5 (~$2.40/task) by about 60%, yet costs nearly twice GLM-5.2 (~$0.47/task). Teams chasing lowest per-task cost should compare both numbers, not just the headline SWE-bench row.

What KDA, AttnRes, and the three infra releases actually do

Moonshot rebuilt three default deep-learning components — attention, residual connections, and the optimizer — rather than scaling an existing recipe.

Kimi Delta Attention (KDA) replaces Gated DeltaNet's scalar forgetting gate with channel-wise gating: each feature dimension gets its own decay rate, implemented as a chunkwise Diagonal-Plus-Low-Rank recurrence. K3 interleaves KDA with Gated MLA global-attention layers, which is how it holds a 1M-token window while keeping KV cache size bounded.

Attention Residuals (AttnRes) swap uniform layer accumulation for input-dependent aggregation across preceding layers. Overhead is one RMSNorm and one pseudo-query vector per layer; Moonshot reports roughly 25% training efficiency gain for under 2% added parameter cost.

Per-Head Muon extends the Muon optimizer to operate per attention head instead of treating the full model uniformly — a training-time technique invisible at the API layer, but part of why K3 punches above its active-parameter count.

Six steps to evaluate Kimi K3 after the July 27 weight release

Headline specs and political posts are not a deployment plan. Run the same checklist whether you call the API or attempt a weight download — a clean macOS box with root access keeps logs reproducible.

1. Read the license before you ship: Confirm your use case clears both gates — no MaaS business above $20M trailing revenue without a Moonshot agreement, and UI attribution if you exceed 100M MAU or $20M monthly revenue. Most startups pass both checks; resellers do not.
2. Choose API vs self-host realistically: If you lack a 64+ GPU supernode, skip the 1.56TB download and route through Moonshot's API or a gateway — our OpenRouter setup guide covers multi-provider routing with fallback chains.

from openai import OpenAI
client = OpenAI(api_key="...", base_url="https://api.moonshot.ai/v1")
resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain KDA in three bullets."}],
    extra_body={"reasoning_effort": "max"},
)
print(resp.choices[0].message.content)

3. Verify cache-hit pricing on your workload: Moonshot's Mooncake disaggregated serving reportedly holds above 90% cache hit rates on coding tasks. Log input token counts and billed amounts — most real calls should cluster near $0.30/M input, not the $3.00 cache-miss headline.
4. Reproduce one benchmark claim locally: Pick a SWE-bench-style coding task from your own repo, run it on K3 and your incumbent model with identical token caps, and compare pass rate and wall time — not just vendor tables.
5. Inspect the weight bundle if self-hosting: Confirm MXFP4 weight + MXFP8 activation format, expert count (896 routed, 16 active per token), and that your serving stack supports sparse MoE at this scale before provisioning hardware.
6. Document provenance for compliance: Record which model version answered which workflow, what data crossed each API boundary, and whether geopolitical policy shifts require a fallback route — especially if your org tracks US-China AI vendor risk.

Numbers worth citing

• Scale: 2.8T total parameters, ~104B active per token, 896 routed experts with 16 activated (~1.8% sparsity), plus shared experts for baseline stability.
• Context and format: 1,000,000-token window; MXFP4 weights with MXFP8 activations; ~1.56TB download on Hugging Face.
• SWE-bench Verified (Vals AI, July 2026): K3 at 93.4% — behind Claude Opus 5 (97%), GPT-5.6 Sol (96.2%), and Claude Fable 5 (95%), ahead of Qwen3.7-Max (79.4%) and DeepSeek-V4 (76.2%).
• Artificial Analysis Intelligence Index (max reasoning): K3 ~57, ranked third overall and first among open-weight models; GLM-5.2 prior leader at 51.
• API pricing: $3.00/M cache-miss input, $0.30/M cache-hit input, $15.00/M output (including reasoning trace).
• License gates: $20M trailing-12-month MaaS revenue threshold; 100M MAU or $20M monthly revenue attribution requirement.
• Self-host floor: Moonshot recommends 64+ accelerator supernode configurations.
• Release timeline: API launch July 16 → White House distillation posts July 22–23 → full weights July 27 → China Commerce Ministry response July 28.

Release details, pricing tiers, and legal allegations can change after publication — treat the primary sources below as authoritative over this article.

Official Moonshot AI materials for the Kimi K3 open-weight release:

Moonshot AI — Kimi K3: Open Frontier Intelligence

Open-source infrastructure accompanying the release:

Moonshot AI — FlashKDA on GitHub

Third-party reporting on the US-China distillation dispute context:

TechCrunch — Researchers skeptical of White House distillation timeline claims

FAQ

The practical limit is the same whether you download weights or call the API: you cannot settle "worth it" from a trending badge alone. K3 rewards teams that need the strongest open-weight capability with a 1M-token window — but the license gates, 64+ GPU self-host floor, and distillation cloud mean you should log identity probes, read the custom license, and keep a compliance-grade fallback. That kind of side-by-side evaluation needs a machine you control end to end, not a shared sandbox with stale Python stacks. A freshly provisioned KVMFLUX Mac mini M4 gives you root access, SSH in minutes, and a wipe-and-retry path when an API key or model ID changes overnight.