Qwen3-TTS Fast

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 62, 65, 64 and 63. Both sweep pairs missed the publish gate on unstable anchors (cartesia:sonic-3.5 and deepgram:aura-2), and smallest:lightning_v3.1 failed over on every call. This row stayed within 5%.

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 point two millions in the seed round.”

• fatal · large numbers — never said “four thousand ninety six” (7 of 7 transcripts)

→ we sent “The file is 4096 bytes.”

→ it said “The file is four hundred ninety-six bytes.”

• fatal · math operators — never said “three quarters” (6 of 7 transcripts)

→ we sent “Give me 3/4 of the batch.”

→ it said “Give me March fourth of the batch.”

+ 13 more fined

Vendor list, Partner tier (docs.narilabs.com/models-and-pricing, read 2026-09-13). Nari also lists a `qwen3-tts-fast:free` tier at $0; this row measured the PAID `qwen3-tts-fast` id, so the paid rate is the one published. The standard (non-fast) qwen3-tts is $5/1M.