Nari Labs Launches Qwen3-ASR Streaming STT: 40ms Fast, $0.06/hr Standard

Nari Labs, a Y Combinator-backed voice AI startup, announced a streaming Speech-to-Text (STT) service on September 12, 2026 KST using Qwen3-ASR 1.7B on its inference engine. One distinction matters: the 40ms latency figure and the $0.06/hour price do not describe the same endpoint. Nari Labs' current product page lists separate Fast and Standard variants.

Image source: Toby Kim (@doyeob) via X

Fast is roughly 40ms at $0.12/hr; Standard is $0.06/hr

Nari Labs' official Model APIs page currently lists the Qwen3-ASR 1.7B streaming STT lineup as follows:

• Nari Qwen3-ASR 1.7B Fast: roughly 40ms server-side TTFS, with listed paid pricing of $0.12 per audio hour.
• Nari Qwen3-ASR 1.7B Standard: listed paid pricing of $0.06 per audio hour.
• Public beta: both Fast and Standard are currently free for a limited time.

The launch post similarly promotes “the world's fastest” STT endpoint at 40ms TTFS and “the world's cheapest” streaming STT endpoint at $0.06/hour as separate claims. Combining those two lines into a single product specification mixes Fast latency with Standard pricing.

Benchmark figures published by Nari Labs

In a follow-up post, Toby Kim said that on Coval's September 10, 2026 benchmark snapshot, Nari Qwen3-ASR Fast ranked first in latency at roughly 43ms p50 and third in word error rate at 3.3% among publicly available APIs. Nari Labs' product page summarizes the Fast endpoint at 40ms TTFS.

Nari Labs also claimed in its launch post that its speed result was around three times faster than ElevenLabs Scribe V2, while the $0.06/hour pricing was around nine times cheaper than Gemini Transcribe 3.5. Those are company-published comparisons, and real-world latency and cost can vary by region, network conditions, and workload.

What to watch during the public beta

Early replies in the launch thread included an anecdotal report of poor Japanese recognition and another complaint that the web playground felt slow. These are individual user reports rather than representative service-wide benchmarks.

The product distinction is clear in the current official material: the low-latency Fast endpoint and the lowest-price Standard endpoint are separate variants, and both are currently available in the free public beta.

Sources

• Toby Kim (@doyeob) on X: Qwen3-ASR streaming STT launch announcement
• Nari Labs Model APIs: Qwen3-ASR Fast and Standard specifications and pricing