aliteq.

How much VRAM Kimi K3 (MoE) needs, and what it costs to run

Some links on this page are referral links: we may earn a commission, at no cost to you.

Trying to decide what to buy? The best GPU for Kimi K3 (MoE) →

2779.9B paramsMoE · 896 experts, 16 active1024k max contextkimi-k3-license

Just want to use it? Per-token API

$3.00in·$15.00out

per 1M tokens · Kimi K3 on Novita AI · read Oct 29, 2026

No GPU to manage: you pay per token, which suits occasional use. Your prompts run on Novita's servers, not your own machine.

Price as listed on Novita's pricing page on Oct 29, 2026; it may have changed.

Memory needed, by quantisation

At 16k context. Weights + KV cache + ~0.8 GB overhead. Kimi K3 (MoE) ships in MXFP4, so its own release leads and no bigger copy is listed: it would only store the same numbers in more bytes. Memory is in binary GB (1,024³ bytes), the unit GPU memory is sold in; Hugging Face and Ollama list downloads in decimal GB, about 7% more.

QuantWeightsKV cacheTotalQuality
MXFP4official1453.7 GB0.4 GB1454.9 GBThe model's own release (MXFP4, a 1560.9 GB download): the best quality there is. A bigger copy only stores the same numbers in more bytes. Its 4-bit GGUF builds keep the 4-bit experts as they are, so they come out about this size too.

How these numbers are produced

  • Model shape comes from the model's own config.json — 24 layers, 96 attention heads, 1 KV heads.
  • Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
  • KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
  • Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
  • This is a mixture-of-experts model. Generation reads only the experts active for each token (16 of 896), so throughput is scored on ~104.0B active parameters, not the full 2780B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Hybrid attention: 93 layers = 69 KDA (Kimi Delta Attention, linear: a fixed-size state, not a growing cache) + 24 gated MLA full-attention layers (model card). `layers`/kv_heads/head_dim = the 24 MLA layers: kv_lora_rank 512 + qk_rope_head_dim 64 = 576 cached numbers per token per layer, mapped as kv_heads 1 x head_dim 288 x 2 = 576 (same convention as GLM-5.3). KDA state is fixed, about 69 layers x 96 heads x 128x128 = ~0.2 GB per sequence at bf16, not counted. Active = vendor 104B; params = HF safetensors total (~2.78T; card says 2.8T), routed experts ship in MXFP4 (~1.56 TB of files); includes the vision encoder. Release date = first public commit on HF (27 Jul 2026; the repo itself was created 13 Jun). HF config.json + model card read 29 Oct 2026.
  • Rental prices are pulled hourly from provider APIs (last updated Thu, 12 Nov 2026 22:20:09 GMT). Spot/interruptible pricing can change or vanish without notice.

Try it interactively

Drag the context slider and watch the KV cache fill the card

Worth reading before you buy

Other models