DEPLOY

Open-source LLM · Moonshot AI

Kimi K2

1000 B parameters (32 B active per token) · 128K context · Modified MIT · released 2025-07-11 by Moonshot AI.

Kimi K2 needs about 2000 GB at FP16: no single GPU can hold it. Serving requires an 8-GPU HGX baseboard (INT8/INT4) or a full rack (NVL72, TPU pod) at FP16.

Where this model sits

The largest open-weight LLM to date. Only 32B active per token means inference cost tracks a 32B model, but disk + memory footprint is enormous.

What runs it (production serving)

Datacenter GPUs and rack designs sized for Kimi K2. Consumer cards are not practical at this size.

Smallest datacenter GPU that fits (INT4)

Memory at different context lengths

Weights are fixed. The KV cache grows linearly with context length and can rival the weights at full window. Most guides only quote weights, then hand-wave a 15% overhead, which is only true at short context.

ContextFP16 totalINT8 totalINT4 total
4K (short chat)2000 + 15.3 = 2015.3 GB1000 + 7.6 = 1007.6 GB500 + 7.6 = 507.6 GB
32K (typical RAG)2000 + 122.0 = 2122.0 GB1000 + 61.0 = 1061.0 GB500 + 61.0 = 561.0 GB
128K (full window)2000 + 488.0 = 2488.0 GB1000 + 244.0 = 1244.0 GB500 + 244.0 = 744.0 GB

KV bytes per token = 2 × 61 layers × 128 KV heads × 128 head-dim × precision bytes. INT4 rows keep the KV cache at INT8 (typical serving default).

Quality benchmarks

Vendor-reported scores from the release blog or tech report. Every row cites its source.

BenchmarkScoreSourceDate
MMLU (5-shot)89.5Moonshot Kimi K2 tech report2025-07-11
HumanEval85.7Moonshot Kimi K2 tech report2025-07-11
MATH (0-shot)82.5Moonshot Kimi K2 tech report2025-07-11

Speed and cost to serve

What cloud providers currently charge to serve Kimi K2. Tokens per second is what the provider publishes on their own page. Batch size, prompt length, and inference engine can shift these numbers a lot; treat as directional.

ProviderHardwareSpeed$/M in$/M outAs of
Moonshot APIinternal—$0.15$2.502025-08-01

Cheapest option in this list for output-heavy workloads: Moonshot API at $2.50/M out. Fastest published: not enough data.

License and variants

License
Modified MIT
The gotcha
Modified MIT requires attribution of 'Kimi K2' when total product MAU exceeds 100M or revenue exceeds $20M/mo.
Variants
Base · Instruct
Context window
131,072 tokens (128K)
Architecture
MoE (1T total, 32B active per token)

Compare with

Model card: https://huggingface.co/moonshotai/Kimi-K2-Instruct · Release: Moonshot AI announcement

See also: every open-source LLM we track · all AI chips · rack designs.