Kimi K3
Moonshot AI's 2.8 trillion parameter Mixture-of-Experts model, the largest open-weight model ever released, built for agentic coding, web research and frontend work across a 1M-token context.
Model Story
Kimi K3 is the successor to Kimi K2.6 from Moonshot AI, the Beijing-based lab behind the Kimi family. Moonshot launched K3 in mid-July 2026, API first, and published the weights on Hugging Face and GitHub on July 27, 2026. At 2.8 trillion total parameters it is the largest open-weight model ever released. Each token is routed to 16 of 896 experts, so about 104 billion parameters are active per token, and the context window is 1,048,576 tokens.
The launch numbers were strong. Kimi K3 took the #1 spot on LMArena's Frontend Code Arena with a score of 1,679, and Artificial Analysis placed it #3 on its Intelligence Index v4.1.1 with a 57, comparable to Claude Opus 4.8 and GPT-5.5 and behind Fable 5 and GPT-5.6 Sol. Artificial Analysis has since rescaled its index to v4.3, where every model scores lower: Kimi K3 (max) now sits at 44, just behind MiMo-V2.6-Pro (46) and GLM-5.3 (45) among open-weight models.
The license changed too. Kimi K2 used a Modified MIT license; K3 ships under a custom "Kimi K3 License" with thresholds for very large commercial deployments, including attribution and branding requirements at very large scale and a separate agreement for model-as-a-service resale above roughly $20M in revenue. Most organizations will sit well below those thresholds, but legal teams should read it before deploying.
For Faraday Machines customers, the honest summary is that Kimi K3 is a large-cluster model. It is available on larger Faraday clusters for organizations that need it, while most teams will get better value from the models we ship on our standard tiers.
Key Specifications
| Developer | Moonshot AI |
|---|---|
| Release Date | Mid-July 2026 (API); weights published July 27, 2026 |
| Architecture | Sparse Mixture-of-Experts (MoE), 16 of 896 experts per token |
| Total Parameters | 2.8 trillion |
| Active Parameters | ~104 billion per token |
| Context Window | 1,048,576 tokens |
| License | Kimi K3 License (custom, with thresholds for very large commercial deployments) |
| Weights | Hugging Face and GitHub |
| Size at 4-bit | Roughly 1.5 TB |
| Faraday Fit | Large cluster only (roughly 12+ Mac Studios) |
API Pricing
Moonshot AI API pricing as of October 2026. Cached input reads cost $0.30 per 1M tokens, and Moonshot charges the same rate across the full 1M-token window. That is a big step up from Kimi K2.6's $0.60 input / $2.50 output. On-premises deployment removes per-token costs, but for K3 the hardware is a large cluster (see below).
Benchmarks
The top row is vendor-reported by Moonshot AI; Moonshot also reports 67.5 on DeepSWE. The bottom row comes from independent evaluators. Artificial Analysis Intelligence Index v4.3 (announced September 2026, now v4.3.2) added harder agentic tests, so its scores are not comparable with v4.1 figures such as K3's launch score of 57. On v4.3.2 the best closed model, Claude Opus 5.5 (max), scores 58, so the gap between Kimi K3 and the very top is real. Kimi K3 had no standardized SWE-bench Pro entry as of late September 2026.
On-Premises Deployment
Kimi K3 does not fit Faraday's Starter, Growth or Scale tiers. At 4-bit quantization its weights alone are roughly 1.5 TB, while our largest standard tier, Scale, provides 384GB of unified memory across three Mac Studios. Running K3 on-premises takes a large cluster of roughly 12 or more units, before leaving room for long-context working memory. Moonshot itself recommends supernode setups of at least 64 accelerators for self-hosting. Faraday's pricing calculator covers clusters from 4 to 64 units, so Kimi K3 is available on larger clusters, sized with you for your workload.
For most teams, the models Faraday ships on standard tiers are the better fit: DeepSeek V4.1 Flash (Scale) for agentic coding and long-context work, GLM-5.3-Flash (Growth) for business intelligence, document analysis and multimodal input, MiniMax M3 (Scale) for long-context multimodal agentic tasks, and Qwen3.8-27B (Starter) for fast, lightweight coding and multilingual work.
Frequently Asked Questions
Can I run Kimi K3 on-premises?
Yes, but only on a large cluster. At 4-bit quantization Kimi K3's 2.8 trillion parameters need roughly 1.5 TB of memory, far more than the 384GB in Faraday Machines' three-unit Scale tier. Running it takes a cluster of roughly 12 or more Mac Studios, and Moonshot AI itself recommends supernodes of at least 64 accelerators for self-hosting. Faraday Machines sizes larger Kimi K3 clusters on request.
How much does the Kimi K3 API cost?
Moonshot AI charges $3.00 per 1M input tokens, $15.00 per 1M output tokens and $0.30 per 1M cached input reads for Kimi K3, at the same rate across the full 1M-token window, as of October 2026. That is a big step up from Kimi K2.6's $0.60 input and $2.50 output. Running Kimi K3 on-premises removes per-token costs, but the hardware is a large multi-unit cluster.
Is Kimi K3 open source, and what license does it use?
Kimi K3 is open-weight but not under a standard open-source license. Moonshot AI published the weights on Hugging Face and GitHub on July 27, 2026 under a custom Kimi K3 License with thresholds for very large commercial deployments, including attribution and branding requirements at very large scale and a separate agreement for model-as-a-service resale above roughly $20M in revenue. Kimi K2 used a Modified MIT license.
How does Kimi K3 compare to Claude Opus 5.5?
Kimi K3 trails Claude Opus 5.5 on independent measures: Artificial Analysis scores Kimi K3 (max) at 44 on its Intelligence Index v4.3.2, against 58 for Claude Opus 5.5 at max effort, the top score. Kimi K3 costs about $2.00 per index task versus $5.98 for Opus 5.5 at max effort, and it launched at #1 on LMArena's Frontend Code Arena with a score of 1,679.
References
[1] VentureBeat. (2026). "China's Moonshot AI releases Kimi K3, the largest open-source model ever, rivaling top U.S. systems." July 2026. Available at: venturebeat.com
[2] Artificial Analysis. (2026). "Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index, comparable to Opus 4.8 and GPT-5.5." July 2026. Available at: artificialanalysis.ai
[3] Tom's Hardware. (2026). "Moonshot releases 2.8 trillion parameter Kimi K3." July 2026. Available at: tomshardware.com
[4] OpenRouter. (2026). "Kimi K3: open source insights." Available at: openrouter.ai
[5] Artificial Analysis. (2026). "Artificial Analysis Intelligence Index v4.3." September 2026. Available at: artificialanalysis.ai
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.
Plan a Kimi K3 Cluster, or Start With a Model That Fits
Faraday Machines sizes large clusters for Kimi K3 and ships DeepSeek V4.1 Flash, GLM-5.3-Flash, MiniMax M3 and Qwen3.8-27B on standard tiers. No per-token charges, full data sovereignty, and hardware you own.
Talk to Us About Cluster Sizing