All models › Compare

DeepSeek-V4-Pro (High, 20260813) vs kimi-k2.5-thinking

DeepSeek-V4-Pro (High, 20260813) and kimi-k2.5-thinking are about even on Arena (1451 vs 1445): in a blind vote each wins about half the time.

DeepSeek-V4-Pro (High, 20260813)1451Arena chat · #49 of 416
kimi-k2.5-thinking1445Arena chat · #62 of 416

Arena scores by category

CategoryDeepSeek-V4-Pro (High, 20260813)kimi-k2.5-thinkingAhead
Chat14511445about even

Scores from blind human votes on Arena. A gap under 10 points is too small to call.

Size and where to use them free

DeepSeek-V4-Pro (High, 20260813)kimi-k2.5-thinking
Made bydeepseekmoonshot
Size1598.8 B (75.6 B active)not public
4-bit download970 GB—
Use it freedownload and run at home · MIT licenseModified MIT license

Speed at home

HardwareDeepSeek-V4-Pro (High, 20260813)kimi-k2.5-thinking
RTX 3090does not fitonline only
RTX 4090does not fitonline only
RTX 5090does not fitonline only
RX 7900 XTXdoes not fitonline only
Mac M4 Max 128 GBdoes not fitonline only
Ryzen AI Max+ 395 128 GBdoes not fitonline only

Estimates at 4-bit with an 8K context, 64 GB of system RAM for offload, calibrated on 44 real llama.cpp tests.

Try them yourself

Ask both the same question side by side in Try & compare, or vote blind between free models in Free Arena.

More comparisons

Share this page:XLinkedInRedditWhatsApp