On the evening of July 27, 2026, Moonshot AI released the full weights and technical report for Kimi K3 — a 2.8-trillion-parameter MoE model with a 1-million-token context window — alongside three open infrastructure projects: MoonEP, FlashKDA, and AgentEnv. This is the largest open-weight model ever shipped in full, a 1.56TB download that hit #1 on Hugging Face trending within thirty minutes. For developers evaluating open-weight licensing, MoE architecture, and self-hosting feasibility, this article covers: the 11-day timeline from API launch to full weight release, KDA / AttnRes / Per-Head Muon architecture breakdown, three infra open-source releases, SWE-bench and AA Index benchmarks, two commercial license gates, API pricing and the 64-GPU self-hosting reality, a six-step evaluation checklist, and FAQ. Bottom line: K3 is the capability ceiling of the open-weight tier — Moonshot never calls it "open source," only open weight.
Three questions developers keep asking: the API went live July 16 — what did July 27 actually add? What's the difference between "open source" and "open weight"? And how many GPUs does self-hosting really need? The timeline below answers all three.
| Date | Event |
|---|---|
| July 16 | Kimi K3 launches on kimi.com, Kimi Work, Kimi Code, and the API — weights not yet public |
| July 17 | Industry analysis intensifies; Xinhua reports it as "the world's largest open model by parameter count" |
| July 22–23 | US accuses Moonshot of "industrial distillation" against Anthropic's Fable model; sanctions threatened |
| July 27, 23:00 | Full weights and technical report released; MoonEP and AgentEnv open-sourced (FlashKDA was already out) |
| July 28 | China's Ministry of Commerce responds to the US-China AI dispute; domestic media covers open-weight details |
License misread risk: Press coverage says "open source," but Moonshot never uses that term — the custom license includes MaaS revenue and attribution gates
Self-hosting illusion: 1.56TB weights plus 896-expert MoE cannot run on consumer GPUs — blind hardware purchases are the biggest hidden cost
Benchmark misuse: Vendor-reported and independently re-tested scores diverge widely — prioritize Vals AI and Artificial Analysis data
Cache pricing trap: Headline input is $3/M, but Mooncake architecture hits 90%+ cache hit rates on coding workloads — real cost is closer to $0.30/M
Geopolitical compliance: US-China distillation dispute plus WAIC 2026 timing — public weights do not mean auditable training
Agent runtime environment: Long-session coding agents need stable compute and toolchains — laptops or low-spec VPS instances cannot sustain production pipelines
| Spec | Value |
|---|---|
| Total parameters | 2.8 trillion (2.8T) |
| Active parameters | ~104 billion |
| Architecture | Mixture-of-Experts (MoE), 896 routed experts, 16 activated per token |
| Attention | Kimi Delta Attention (KDA) + Gated Multi-Head Latent Attention (MLA) |
| Context window | 1,000,000 tokens |
| Multimodality | Native vision understanding (ViT-V2, 27 layers) |
| Weight format | MXFP4 weights + MXFP8 activations (quantization-aware training) |
| Download size | ~1.56TB (Hugging Face) |
| License | Custom license — Moonshot calls it "open weight," not "open source" |
Standard Gated DeltaNet applies a single scalar forgetting gate to the entire memory state. KDA replaces that with channel-wise gating — every feature dimension gets its own independent decay rate. Implemented as a chunkwise Diagonal-Plus-Low-Rank (DPLR) formulation, it keeps recurrence linear in time while staying hardware-efficient. K3 interleaves KDA layers with a smaller number of Gated MLA global-attention layers — the hybrid design that supports a 1M-token context while keeping KV cache size low.
Standard residual connections accumulate every layer's output uniformly as depth increases — early-layer information gets diluted in deep networks. AttnRes replaces uniform accumulation with selective, input-dependent aggregation across all preceding layers. Overhead is minimal: one RMSNorm and one pseudo-query vector per layer. In testing, this delivered roughly 25% higher training efficiency for under 2% added parameter cost.
Per-Head Muon extends the Muon optimizer to operate per attention head rather than treating the whole model uniformly. The MoE layer routes across 896 experts and activates 16 per token (~1.8% sparsity), with Quantile Balancing and MoonEP proving a mathematical upper bound on redundant experts per compute rank — solving large-scale MoE load imbalance.
Moonshot rebuilt attention, residual connections, and the optimizer — three of deep learning's oldest defaults. That engineering depth, not raw parameter count, is what separates K3 from a simple scale-up.
This release is not just weights — Moonshot open-sourced the infrastructure stack that made K3 training possible, which is a major reason the developer community responded positively.
| Technology | Role | Key capability |
|---|---|---|
| MoonEP | High-performance communication for large fine-grained MoE | Temporarily duplicates overloaded experts so every rank gets equal token counts; proves redundant-expert upper bound |
| FlashKDA | CUTLASS-based KDA kernel (previously open-sourced) | 1.72–2.22x faster prefill on NVIDIA H20 vs. flash-linear-attention baseline; drop-in chunk_kda backend |
| AgentEnv | Firecracker microVM agent sandbox (co-developed with KVCache.ai) | Checkpoint latency as low as 133ms, resume 49ms, up to 6.5x memory overcommit (Moonshot-reported, not independently verified) |
Confirm your path: Most teams should use the API (https://api.moonshot.ai/v1, model ID kimi-k3) — self-host only for data residency, fine-tuning, or extreme inference volume
Read the LICENSE line by line: Check the $20M MaaS revenue gate and the 100M MAU / $20M monthly revenue attribution requirement
Verify Hugging Face weight integrity: ~1.56TB total — confirm MXFP4 format and shard manifest
Assess hardware requirements: Moonshot recommends 64+ accelerator supernodes — do not plan around consumer GPUs
Integrate FlashKDA: Existing flash-linear-attention projects can swap the chunk_kda backend for immediate speedup
Cross-check independent benchmarks: Use Vals AI SWE-bench Verified (K3: 93.4%) and AA Index (~57, #3 globally) — not vendor self-reports alone
| Model | Score | Release date |
|---|---|---|
| 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 |
| GLM-5.2 (prior open-weight leader) | — | AA Index 51 |
| Dimension | Kimi K3 | Claude Fable 5 | GLM-5.2 |
|---|---|---|---|
| AA Intelligence Index | ~57 (#3 globally, #1 open-weight) | 60 | 51 |
| Per-task cost | ~$0.95 | ~$2.40 | ~$0.47 |
| Context | 1M tokens | Shorter | Shorter |
| Weights downloadable | Yes (1.56TB) | No | Yes |
| Positioning | Open-weight capability ceiling | Closed-source top tier | Best open-weight value |
Two commercial gates (new in K3, absent from K2): (1) Operating a MaaS business with trailing-12-month revenue over $20 million requires a separate commercial agreement with Moonshot; (2) Products exceeding 100 million MAU or $20 million monthly revenue must prominently display "Kimi K3" in the UI. Irrelevant for most teams — but if you plan to compete directly with kimi.com's API, legal review is mandatory.
| Token type | Price |
|---|---|
| Input (cache hit) | $0.30 |
| Input (cache miss) | $3.00 |
| Output (including reasoning trace) | $15.00 |
Kimi K3's open-weight release landed during WAIC 2026 and days after the US-China "model distillation" dispute escalated: White House science adviser Michael Kratsios accused Moonshot of running a "large-scale, covert industrial distillation" campaign against Anthropic's Fable model; Treasury Secretary Scott Bessent floated sanctions and Entity List designation. China's Ministry of Commerce pushed back on July 28, accusing Washington of "AI hegemonism." Releasing full weights, the technical report, and three infrastructure projects reads as a deliberate signal — engineering transparency as the clearest answer to questions about training methods. Three days later, Alibaba (a Moonshot investor) unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter model widely read as a direct response — pushing the open-weight race into the "3T club."
If you plan to wire K3 into Kimi Code-style long-session agents or iOS CI pipelines, running CLI agents on a laptop or unstable Linux VPS often means out-of-memory crashes, dropped sessions, and missing Xcode/Metal toolchains. For production environments that need stable SSH sessions, DerivedData caching, and iOS CI/CD automation, NodeMini's cloud Mac Mini rental is usually the better fit — dedicated nodes with second-scale provisioning let agents and build jobs run continuously on a real Mac. See Mac Mini rental rates.
No, not by the OSI's strict definition. It is open-weight: weights, the technical report, and some infra tools are public, but training data and full training code are not. Moonshot's own materials never use the term "open source."
Yes, for the vast majority of use cases. Restrictions apply only if you run a MaaS business generating over $20M in trailing-12-month revenue, or if your product exceeds 100M MAU or $20M monthly revenue. For deployment infrastructure, see Mac Mini rental rates.
Moonshot recommends a 64+ accelerator supernode cluster. For most developers and companies, the practical path is the official API or OpenRouter and similar platforms (seven providers already host K3).
Strongest in the open-weight tier: AA Index #3 globally, behind only Claude Fable 5 and GPT-5.6 Sol — but more expensive than GLM-5.2. Highest capability ceiling, not the best value pick. More operational questions in the help center.
K3 has roughly 3x the parameters of K2.5, adds Attention Residuals and Per-Head Muon, and significantly expands context and multimodal capability. The license moved from K2's Modified MIT to a custom file with a new MaaS revenue threshold.