What Do You Need to Run Kimi Locally? VRAM by Version
Kimi currently has 6 versions you can deploy locally. The smallest, Kimi Linear 48B A3B, needs 32GB of VRAM at 4-bit; the largest, Kimi K3, is an MoE model and needs 32GB of VRAM plus about 1,480GB of RAM with MoE offloading. The cheapest device that gets one running is the Tesla V100 32GB, at about $895.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| Kimi K3MoE | Jul 27, 2026 | 2,800B | 104B | 1M | Kimi K3 License |
| Kimi K2.7 CodeMoE | Jun 12, 2026 | 1,000B | 32B | 256K | Modified MIT |
| Kimi K2.6MoE | Apr 20, 2026 | 1,000B | 32B | 256K | Modified MIT |
| Kimi K2.5MoE | Jan 27, 2026 | 1,000B | 32B | 256K | Modified MIT |
| Kimi Linear 48B A3BMoE | Oct 30, 2025 | 49.1B | 3B | 1M | MIT |
| Kimi K2MoE | Sep 3, 2025 | 1,000B | 32B | 256K | Modified MIT |
About Kimi
Kimi is the large model family from Moonshot AI. Yang Zhilin and others founded Moonshot AI in Beijing in March 2023, and it launched the Kimi assistant that October. Since 2025 it has been releasing model weights one after another. Its flagships are all very large MoE models aimed at agents, coding and long context.
Kimi K2, from July 2025, has 1 trillion parameters with 32 billion active per token, and uses MLA attention and Moonshot's own MuonClip optimizer; the reasoning version, K2 Thinking, followed in November. Kimi Linear, released in October 2025, introduced the company's own linear attention, KDA. K2.5, in January 2026, became natively multimodal, followed by K2.6 and the coding-focused K2.7 Code. K3, in July, scaled up to 2.8 trillion parameters with 104 billion active per token, mixes KDA and MLA, and has a 1-million-token context; at release it was the open-weight model with the most parameters. On licensing, the K2 series uses a modified MIT license that allows commercial use but requires very large products to display the model name. K3 adds a rule that model service providers with annual revenue above 20 million US dollars must sign a separate agreement before commercial use. Also, Cursor's coding model Composer 2 was built on Kimi K2.5.
- VendorMoonshot AI
- Versions6
- Parameters49.1B – 2,800B
- Latest versionKimi K3
- License
- MIT / Modified MIT / Kimi K3 License
- Official site
- www.moonshot.ai
- Hugging Face org
- moonshotai
How much VRAM do Kimi models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a Kimi version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Kimi K32,800B (104B active) | 861.3GGUF | 863.3 | 866 | 869.6 | 891.4 |
| Kimi K2.7 Code1,000B (32B active) | 339.5GGUF | 342.9 | 349.8 | 359 | — |
| Kimi K2.61,000B (32B active) | 340.5GGUF | 343.9 | 350.8 | 360 | — |
| Kimi K2.51,000B (32B active) | 374.8GGUF | 378.2 | 385.1 | 394.3 | — |
| Kimi Linear 48B A3B49.1B (3B active) | 17.9GGUF | 19.2 | 20 | 21.1 | 27.4 |
| Kimi K21,000B (32B active) | 383.6GGUF | 387 | 393.9 | 403.1 | — |
How much VRAM do Kimi models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a Kimi version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Kimi K32,800B (104B active) | 1,508.7GGUF | 1,510.7 | 1,513.4 | 1,517 | 1,538.8 |
| Kimi K2.7 Code1,000B (32B active) | 583.7GGUF | 587.1 | 594 | 603.2 | — |
| Kimi K2.61,000B (32B active) | 583.7GGUF | 587.1 | 594 | 603.2 | — |
| Kimi K2.51,000B (32B active) | 621.2GGUF | 624.6 | 631.5 | 640.7 | — |
| Kimi Linear 48B A3B49.1B (3B active) | 30.1GGUF | 31.4 | 32.2 | 33.3 | 39.6 |
| Kimi K21,000B (32B active) | 621.3GGUF | 624.7 | 631.6 | 640.8 | — |
How much VRAM do Kimi models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a Kimi version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Kimi K32,800B (104B active) | 1,561.2GGUF | 1,563.2 | 1,565.9 | 1,569.5 | 1,591.3 |
| Kimi K2.7 Code1,000B (32B active) | 594.5GGUF | 597.9 | 604.8 | 614 | — |
| Kimi K2.61,000B (32B active) | 594.5GGUF | 597.9 | 604.8 | 614 | — |
| Kimi K2.51,000B (32B active) | 1,091.1GGUF | 1,094.5 | 1,101.4 | 1,110.6 | — |
| Kimi Linear 48B A3B49.1B (3B active) | 52.2GGUF | 53.5 | 54.3 | 55.4 | 61.7 |
| Kimi K21,000B (32B active) | 1,091.1GGUF | 1,094.5 | 1,101.4 | 1,110.6 | — |
FAQ
Is Kimi open source and free for commercial use?
It depends on the version; there are three licenses in all. Kimi Linear 48B A3B uses the standard MIT license. K2, K2.5, K2.6 and K2.7 Code use Moonshot AI's modified MIT license, which adds one clause: if you use the model in a commercial product with more than 100 million monthly active users or monthly revenue above a set threshold, the product's interface must prominently display the corresponding model name, such as “Kimi K2.6.” K3 moved to a separate Kimi K3 License. On top of the same attribution requirement, it says that a corporate group running a “model as a service” business (offering inference or fine-tuning to third parties through an API or similar means), and whose revenue over 12 consecutive months exceeds the threshold, must sign a separate agreement with Moonshot AI before using the model for any commercial purpose. The K3 license also states that neither clause applies if you use the model only internally and don't offer its capabilities to third parties.
Can you run Kimi with Ollama?
Not locally. On Ollama, Kimi K2.6, K2.7 Code and K3 have only cloud tags (kimi-k2.6:cloud, kimi-k2.7-code:cloud and kimi-k3:cloud), which run on Ollama's servers. The kimi-k2 and kimi-k2.5 pages have been taken down, and Kimi Linear isn't listed at all. To deploy locally, use llama.cpp with GGUFs: unsloth has made files for K2 through K2.7 Code and for K3, and bartowski has made them for Kimi Linear. Note that in llama.cpp, K3 currently handles text only and can't read images. You can also run the official weights on vLLM.
