Qwen3-TTS

Deterministic: identical audio every run. The endpoint pins seed 0 (x-nari-seed) on a sampling model, so this describes the deployment, not the voice. A response cache was ruled out.

Four n=30 sweeps from US East on 2026-09-26 read p50 74, 76, 76 and 75, in the same runs as qwen3-tts-fast and with the same anchor misses. This row stayed within 3%.

Fined:

• fatal · currency — never said “three point two million” (48 of 49 transcripts)

→ we sent “We raised $3.2M in the seed round.”

→ it said “We raised three dollars lot twenty-one in the seed round.”

• fatal · large numbers — never said “three trillion” (49 of 49 transcripts)

→ we sent “National debt passed 3,000,000,000,000 dollars.”

→ it said “National debt passed $3 billion.”

• fatal · math operators — never said “one thousand twenty four” (43 of 49 transcripts)

→ we sent “Since 2^10=1024, it fits in one block.”

→ it said “Since two to the power of ten equals w times ten twenty or twenty-four, it fits in one blo”

+ 13 more fined

Vendor list $5 per 1M input characters (docs.narilabs.com/models-and-pricing, read 2026-09-26).