Herds by machine count, each machine counted once from its latest report in this window. Self-reported by herds with telemetry on; not verified.
Do bigger herds work harder?
Requests routed per machine in the last 30 days by the typical herd of each size.
The median herd of each size: each herd's requests divided by its machines, idle herds included. A median, so one very busy herd can't move it. Self-reported; not verified.
Models the most herds run
Counted by herds, not requests. One busy herd can send most of the traffic in the request table below, so this is the better read on what is common.
Top 10 models by herds that routed at least one request to them, with "model:latest" counted as "model". Self-reported by herds with telemetry on; not verified.
Most-used models
#
Model
Requests
Share
1
gpt-oss:120b
216,022
69.77%
2
dinov2-vit-s14
37,068
11.97%
3
nomic-embed-text
11,202
3.62%
4
qwen3.5:9b
6,912
2.23%
5
llama3.2:1b
6,187
2.0%
6
qwen3-coder:30b
6,047
1.95%
7
gemma3:12b
4,669
1.51%
8
gemma4:e4b
4,664
1.51%
9
qwen3-embedding
1,981
0.64%
10
qwen3:4b
1,656
0.53%
11
qwen3.8:27b-mlx
1,610
0.52%
12
qwen3.5-4b-sompoa-heresy-v2:q4_k_m
1,587
0.51%
13
llama3.1
1,523
0.49%
14
qwen2.5-coder:14b
1,443
0.47%
15
qwen3-embedding:0.6b
1,443
0.47%
16
qwen3-coder
1,246
0.4%
17
qwen3:32b
967
0.31%
18
gemma4:12b
713
0.23%
19
qwen2.5:7b
570
0.18%
20
gemma4b:31-code-frame
419
0.14%
21
qwen3.8:27b
352
0.11%
22
gemma4:31b
345
0.11%
23
qwen3.6-35b-a3b
340
0.11%
24
gpt-oss:20b
323
0.1%
25
gemma3:27b
313
0.1%
The herd leaderboard
Herds that report in, by name where the operator picked one and by a generated handle otherwise. Ranked by total memory, because that is more fun to argue about than request counts.
Error categories are a fixed list. Raw error text never leaves anyone's machine.
About these numbers
Herds report a daily summary unless the operator turns it off with FLEET_NODE_TELEMETRY=false. Every field is listed on the telemetry page. The two herd counts differ on purpose: an idle herd reports its machines without routing any requests. Everything is self-reported, so read it as a strong signal, not a census.
Machine-readable
The same data is available as JSON at /api/v1/stats.