Qwen3-30B-A3B

Model by Qwen team (Alibaba).

Modeltext in, text out

Input

text

Output

text

Chat with streaming, OpenAI-compatible.

What you send: text
What you get: text

Available as a Pack. Write to [email protected].

Qwen3-30B-A3B at 4-bit on one TPU v6e, 1, 10 and 64 users at once

Users at once	Output tokens a second, all users	p95 wait to first token	Our cost per 1M output tokens
1	301.6	317 ms	$2.05
10	409.3	954 ms	not computed
64	1,906.7	5,011 ms	$0.32

In 28 of 28 runs, up to 256 users at once, every user's tokens matched running alone, bit for bit. Measured on one chip, benchmark, 23 October 2026: a chip test, not live serving. The 1-user row is single-user mode; its wait time is from the batched run. Cost is computed at the TPU v6e on-demand list price.