Eugene Ye

ops & compute @wafer_ai, prev @mercor_ai, @harvard

[wafer @wafer_ai]

bro... we've made it

@karpathy hi btw 🧇

[Eugene Ye @eug0212]

Is memory the moat?

@wafer_ai engineers figured out how to serve Kimi K3 on AMD MI355X at 952 tok/s/node and 118 tok/s single stream.

This is over 3.8x the aggregate throughput/node and over 1.3x the single stream decode of B200.a

This also crushes B300 on performance per dollar: 48 vs 33 tok/s/$.

@ianye0212 wrote about how we did this https://t.co/9uthIKPpMg

[Eugene Ye @eug0212]

https://t.co/LS4sRICeMA

[wafer @wafer_ai]

🚨 BREAKING:

wafer now runs the fastest Kimi K3 in the world

ranked #1 across every provider on @ArtificialAnlys

⚡ 172 output tok/s
⚡ 15.8s end-to-end response time

get this performance on a dedicated endpoint: https://t.co/TsEoUs3wpn

[Eugene Ye @eug0212]

W

[Danny Liu @gpudanny]

holy decode

[wafer @wafer_ai]

🚨 BREAKING

Kimi K3 Fast⚡️is live on @OpenRouter

the fastest serverless Kimi K3 in the world

pick @wafer_ai as your provider: https://t.co/xyXjEEV2v5

[Eugene Ye @eug0212]

👀

[steve @gpusteve]

kimi k3 fast is live on @OpenRouter

fastest serverless kimi k3.

pick @wafer_ai as your provider: https://t.co/0fB588bjrZ

[Eugene Ye @eug0212]

@gpusteve Bro beats me in smash and tweets this immediately after

[steve @gpusteve]

we're shipping kimi k3 FAST next week. fastest serverless kimi k3.

unlock 2x credits for kimi k3 Fast.

follow @wafer_ai and sign up here: https://t.co/rfurdSXFqF

[wafer @wafer_ai]

🚨 BREAKING:

GLM-5.2 Fast (⚡️150-250+ tok/s) is 30%+ off through July 31.

old: $3.00 in / $10.25 out / $0.50 cache
NEW: $2.10 in / $6.60 out / $0.21 cache

live on @wafer_ai, @OpenRouter, and @vercel AI Gateway. pick wafer as your provider.

(screenshot from Vercel AI Gateway)

[Harvard University @Harvard]

Welcome to Harvard Class of 2025! 🎉 #Harvard2025 https://t.co/xcTlM7TVN9