Moonshot's Kimi K2 generation is a 1030B MoE design: K2.6 for general chat and K2.7-Code for coding, both at 256K context under a Modified MIT license. These are datacenter-scale models — far beyond any single consumer GPU.
Kimi is Moonshot AI's open-weight flagship line, and the current K2 generation is a family of two: Kimi-K2.6 for general chat and reasoning, and Kimi-K2.7-Code for programming and agentic coding. Both are community favorites released in 2026, and both share one architecture: a 1030B-parameter MoE with 384 experts, 8 of them active per token — only about 32B parameters fire for each token generated.
That MoE design is what makes a trillion-parameter model servable at all. Per-token compute behaves much more like a 32B dense model than a 1030B one, so tok/s on adequate hardware is surprisingly usable. But "adequate hardware" is doing heavy lifting in that sentence: all 1030B parameters still have to sit in memory somewhere, and that makes Kimi the largest family in our catalog by a wide margin.
So the honest headline up front: these are datacenter-scale models. No single consumer GPU — not even a card in the 192GB+ tier — holds Kimi K2 at Q4. This page tells you exactly how much memory is required, and what to buy instead if your budget is one graphics card.
Where does 574 GB come from? Our fit formula is weights + KV cache + 1 GB overhead. The Q4 weights alone are 566.5 GB, and at the default 8K context the KV cache adds only about 6 GB — because for an MoE model, KV cache scales with the active 32B, not the full 1030B. Context length is the trap, though: stretch toward the full 256K window and the KV cache balloons to roughly 205 GB on its own, so a long-context deployment really wants ~780 GB or more of aggregate VRAM at Q4.
Lower quants help only so much. At Q2 the weights drop to 288.4 GB — still far beyond any single card we list. FP16 weights would be about 2,060 GB (~2 TB), firmly multi-node territory; in practice, everyone runs this family quantized.
The first thing is MoE economics. You pay the memory bill for 1030B parameters but the compute bill for 32B, which is why these models can feel quick on big iron while being impossible to fit on a desktop. The second thing is context: both models ship with a 256K window, long enough for whole codebases or book-length documents — provided you budget the extra KV cache, which grows with context length even though the weights do not. The third thing is licensing: Kimi K2 uses a Modified MIT license. It is permissive by the standards of trillion-parameter flagships, but it is not plain MIT — read the actual terms before a commercial deployment.
By VRAM tier, honestly:
Everyone else is better served by a smaller family — a 70B-class dense model or a sub-100B MoE fits one or two consumer cards and covers most local-LLM use cases. Our VRAM sizing guide walks through the math.
| Model | Params (B) | Ctx (K) | Q4 (GB) | License | Fit check |
|---|---|---|---|---|---|
| Kimi-K2.6 | 1030.0 | 256 | 566.5 | Modified MIT | — |
| Model | Params (B) | Ctx (K) | Q4 (GB) | License | Fit check |
|---|---|---|---|---|---|
| Kimi-K2.7-Code | 1030.0 | 256 | 566.5 | Modified MIT | — |