Model reference · open weights

k2-merged-3.5T

LLMs NousResearch Text gen 2 builds Its own licence terms 330 dl/mo

k2-merged-3.5T is an open-weight language model from NousResearch. k2-merged-3.5T-fp8 (FP8) weighs 3472 GB; it needs more than the reference configurations below.

What it is

Released byNousResearch
Released2025-09-12
Parameters3468.1B
VRAM3472 GB for the weights

What it runs on

Memory and cards for k2-merged-3.5T-fp8 (FP8)

3472 GBweights, file size
119 MBcache per 1K tokens
4.2 GBruntime overhead, at least
131,072 tokenscontext max
CardRequests at onceContext maxMemory
8K each32K each
RTX 3060 12 GB … 8× B200 180 GB
16 smaller cards
———
Memory needed at each load
Requests at once8K tokens each32K tokens each
13477 GB3480 GB
53481 GB3496 GB
83484 GB3507 GB
163492 GB3538 GB
323507 GB3601 GB
643538 GB3725 GB

One card, with vLLM's small-card settings.

Builds

Sizes, precisions & builds

BuildParamsPrecisionWeightsSmallest setup
k2-merged-3.5T-bf16 ↗ 3468.1BBF16 6936 GBnone of the reference configurations
k2-merged-3.5T-fp8 (above) ↗ —FP8 3472 GBnone of the reference configurations

From the model card

What NousResearch says about k2-merged-3.5T

Experimental frankenmerge of Kimi K2-07, 09 and Base

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms