Kimi K3 is Moonshot AI’s July 2026 frontier model and, at 2.8 trillion parameters, the world’s first open-weight model in the 3T class. It launched via hosted API on July 16, with full weights and a 47-page technical report following on July 27. The pitch is straightforward: frontier-level intelligence that you can download, fine-tune and self-host — with caveats on both the license and the hardware needed to actually run it.
The architecture is a Mixture-of-Experts transformer with 896 total experts, 16 active per token (104B activated parameters), a 1M-token context window, and native vision via MoonViT-V2. Moonshot claims ~2.5× scaling efficiency over Kimi K2 through three innovations: Kimi Delta Attention (KDA), Attention Residuals (AttnRes) and Stable LatentMoE.
Architecture
| Spec | Value |
|---|---|
| Total parameters | 2.8T |
| Activated parameters | 104B (16 of 896 routed experts per token) |
| Architecture | MoE transformer + KDA + AttnRes + Stable LatentMoE |
| Context window | 1M tokens |
| Vision | MoonViT-V2 (native image understanding) |
| API model ID | kimi-k3 |
| Post-training | RL across general, agentic and coding domains + multiple reasoning-effort levels |
Pricing (hosted API)
| Rate / 1M tokens | |
|---|---|
| Input | $3.00 |
| Cached input | $0.30 |
| Output | $15.00 |
The weights are free to download, so the alternative to the API meter is self-hosting — if you have the compute.
The license question
Kimi K3 does not ship under MIT or Apache 2.0. It uses a bespoke Kimi K3 License with two key conditions:
- MaaS revenue gate: Model-as-a-Service operators whose group revenue exceeds $20M over any consecutive 12 months must sign a separate commercial agreement with Moonshot AI before commercial use.
- Branding mandate: Commercial products above 100M MAU or $20M monthly revenue must display “Kimi K3” prominently in the user interface.
Section 4 exempts purely internal use and access through Moonshot’s own products. For most individual developers and smaller companies, the license is permissive. For enterprises building hosted inference services, read the LICENSE file on the repository carefully.
Benchmarks
Moonshot’s technical report positions K3 between the top open models and the strongest proprietary ones. Per the abstract: K3 “consistently outperforms other open and proprietary models” but “still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol.”
The model was trained with reinforcement learning across general, agentic and coding domains with multiple reasoning-effort levels — enabling compositional generalization and long-horizon execution. Specific benchmark numbers are in the arXiv paper (2607.24653); we will update this page when independent replications land.
Self-hosting
Moonshot recommends deploying K3 on supernode configurations of 64+ accelerators and points to vLLM, SGLang and TokenSpeed as supported inference engines. Key infrastructure details:
- Weights ship natively quantized (quantization-aware training from the SFT stage)
- Moonshot contributed a custom caching implementation to vLLM because KDA breaks conventional prefix caching
- The release includes attention kernels and an MoE communication library
For most teams, the near-term practical value is downstream distillation, fine-tuning and research access to the training recipe — not raw self-hosting of the full 2.8T model.
How it compares
- vs Kimi K2.7 Code: K3 is the much larger successor (2.8T vs 1T). K2.7 Code was coding-focused with Modified MIT; K3 is general-purpose frontier with a more restrictive custom license.
- vs GLM-5.2: Both are large Chinese open-weight MoEs. GLM-5.2 (753B, MIT) is much smaller and truly MIT-licensed; K3 (2.8T, custom license) is larger and more capable, but harder to self-host and more restrictive commercially.
- vs Claude Fable 5 / GPT-5.6 Sol: K3 trails both on the hardest benchmarks per its own disclosure, but is the only one whose weights you can download. For organizations that can self-host, the value proposition is control and long-term cost.
Timeline
| Date | Event |
|---|---|
| July 16, 2026 | Hosted API launch ($3/$15) |
| July 27, 2026 | Full weights + tech report + infrastructure released on GitHub / Hugging Face |
Best for
- Downstream distillation and fine-tuning from frontier-quality open weights
- Self-hosted frontier inference for organizations with the compute
- Long-horizon agentic coding and knowledge work
- Research access to 3T-class training recipes and infrastructure
Pros & cons
Strengths
- Largest open-weight frontier model at 2.8T parameters — outperforms all other open and most proprietary models
- ~2.5× scaling efficiency over Kimi K2 via Kimi Delta Attention and Stable LatentMoE
- Native vision (MoonViT-V2) and 1M-token context window
- Full weights + training infrastructure (attention kernels, MoE comm library) publicly available
Limitations
- Requires 64+ accelerator supernode — not practical for most teams to self-host
- Custom Kimi K3 License, not MIT/Apache — commercial MaaS operators >$20M revenue need a separate agreement
- Still trails Fable 5 and GPT-5.6 Sol on the hardest benchmarks
- No audio/video input
Frequently asked questions
When was Kimi K3 released?
Moonshot AI launched Kimi K3 via its hosted API on July 16, 2026. The full model weights, 47-page technical report and infrastructure code were released on July 27, 2026 on Hugging Face and GitHub.
Is Kimi K3 truly open source?
Open weights, but not open source by the OSI definition. The weights ship under a custom "Kimi K3 License" that allows free use for most developers but requires a separate commercial agreement for Model-as-a-Service providers with over $20M in annual aggregate revenue, and mandates Kimi K3 branding for commercial products exceeding 100M MAU or $20M monthly revenue. Read the LICENSE file on GitHub directly.
How big is Kimi K3?
2.8 trillion total parameters in a Mixture-of-Experts configuration with 896 total experts and 16 active per token, yielding 104 billion activated parameters. It uses Kimi Delta Attention (KDA), Attention Residuals (AttnRes) and Stable LatentMoE.
What hardware do I need to run Kimi K3?
Moonshot recommends 64+ accelerator supernode configurations. Supported inference engines include vLLM, SGLang and TokenSpeed. The weights ship natively quantized (quantization-aware training from the SFT stage). Moonshot also contributed a custom caching implementation to vLLM because Kimi Delta Attention breaks conventional prefix caching.
How does Kimi K3 compare to closed models?
Per the technical report, K3 "consistently outperforms other open and proprietary models" in their evaluation suite but "still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol." On SWE-bench Pro and Terminal-Bench, K3 sits between Opus 4.8 and Fable 5 levels.