Numes retweeted
a minute of agent traffic on Numes, drawn as a Command & Conquer battle
every unit is a request, every barracks a model, the GPUs are the fortress
live: numes.schnitzel.world/?ref=x…
Numes retweeted
what agents pick on our Numes endpoint, requests since Nov 3:
GLM 5.3 @Zai_org: 618k
Kimi @Kimi_Moonshot: 495k
DeepSeek V4.1 Flash @deepseek_ai: 285k
uncensored builds included. GLM has the most requests, DeepSeek the longest prompts
Numes retweeted
yesterday was our biggest day on Numes so far: 368k requests and 38B tokens, almost double the day before
11 days in: 1.40M requests and 157B tokens
Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, free right now: numes.ai
Numes retweeted
free Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash in @opencode, two commands:
curl -fsSL numes.ai/install | sh
numes run opencode
same for Pi from @earendilworks: numes run pi
the free models run on our GPUs at amazee.ai
Kimi K3.1 Pro by @amazeeai is available for free on Numes.
And now faster than ever before! 🔥
Kimi K3.1 Pro on Numes was too slow last week: 4-8 tok/s and 9-29 s to the first token
last 24h: 26 tok/s and 3 s
Kimi got more GPUs, and we retired models few people used. Token Factory moves GPUs between models by itself
1
8
1,380
Numes retweeted
the median request we get through Numes: 70,497 tokens in, 182 out
38% of requests are over 100k tokens
that's coding agents re-reading their context every turn. a chat app looks nothing like this











