gpurank
Compare pricesCompare
esc
  • Nothing named “”. Press Enter to search the offers instead.
↑↓ to move · ↵ to open · esc to close0 entries
Moonshot AI · Chat and reasoning · 2.8T parameters · 104B active · Kimi K3 License

Kimi K3

Needs 1,874 GB at INT4 / MXFP4. The smallest real configuration today is 8× MI325X at TensorWave, $18.00 per hour, about $13,140 a month at 730 hours.

Parameters2.8T104B active per token
Context1M tokensmaximum published
Checkpoint1,560 GBINT4-packed, as published
LicenseKimi K3 Licensecheck before commercial use
ReleasedJul 16, 2026updated Jul 27, 2026
Downloads854,823moonshotai/Kimi-K3 ↗
Smallest configuration · INT4 / MXFP4
8× MI325X2048 GB VRAM · 91% used
Per hour
$18.00
Per month
$13,140
Cheapest at
TensorWave
All MI325X offers →

What is documented about Kimi K3

news checked 2026-10-27

Announced 2026-07-16; weights on Hugging Face since 2026-07-27 (about 1.56 TB in 96 shards). MoE, multimodal, 1M-token context. Licence clause: any company above $20M annual revenue must negotiate with Moonshot before serving the model to third parties. Moonshot recommends supernodes of 64 accelerators or more, MXFP4 weights with MXFP8 activations. Its successor Kimi K3.1 Pro (2026-10-23) is API-only until the weights promised for 2026-11-04.

  • GPUBank global report, 2026-10-25, §3.4 (internal, sourced)
  • https://huggingface.co/moonshotai/Kimi-K3
  • https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json
  • https://huggingface.co/api/models/moonshotai/Kimi-K3?blobs=true
Memory · 1 precision · 6 context lengths

How much VRAM it needs

Weights plus runtime overhead, then the KV cache that grows with the context you serve.

PrecisionWeightsWeights + headroom4k context32k context128k context256k context1M context
INT4-packedpublished checkpoint1,561 GB1,874 GB1,874 GB1,877 GB1,888 GB1,902 GB1,986 GB
KV cacheone sequence, FP16 cache, included above——1 GB4 GB15 GB29 GB113 GB
Cross-checked on Hugging Face Oct 28, 2026: 2779.9B parameters in the safetensors headers, 1560.9 GB of weight files, attention layout from config.json (93 layers, latent rank 512). Method: weights as published at the native precision, verified on Hugging Face 2026-10-28; other precisions estimated at 1 (8-bit) and 0.6 (4-bit) bytes per parameter; × 1.2 runtime headroom; KV cache = 1152 bytes per token per layer × 93 full-attention layers (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 110 GPUs · cheapest eligible on-demand offer per GPU count, every provider on file

What it costs to run, GPU by GPU

The fewest GPUs of one type whose total memory fits the model at short context, priced from the cheapest eligible on-demand offer for exactly that count (45-day window, newest reading per provider), with the next providers beside it. Monthly = 730 hours.

GPUSmallest configVRAM used$ / hour$ / monthCheapest at
MI325XBest fit8× 256 GB2048 GB total91% tight$18.00$13,140TensorWaveTrust 64 · read Nov 1 · receipt · also Cyfuture AI $24.34
MI355X8× 288 GB2304 GB total81%$23.60$17,228TensorWaveTrust 64 · read Nov 1 · receipt
B3008× 288 GB2304 GB total81%$46.59$34,009UpCloudTrust 75 · read Nov 1 · receipt · also Limestone Networks $50.00, Massed Compute $52.80
B300 CC8× 288 GB2304 GB total81%$92.62$67,613VerdaTrust 81 · read Nov 1 · receipt
Kimi K3 · related
Same publisher or category