Skip to content

Qwen3-ASR 1.7B MLX q8

Current status and review date

Decision snapshot

  • Product role: English STT challenger.
  • Selected or best-qualified configuration: none qualified; exact tested configurations are documented below.
  • Measured hardware: Apple M4 Max, 40-core GPU, 48 GiB unified memory; direct same-host component HTTP.
  • Evidence: functional and matched corpus quality with paired uncertainty.
  • Decision: no-promotion; keep existing controls without newly qualifying them.
  • Important limitation: Critical entity error; paired uncertainty overlaps zero; microphone/client gates missing.
  • Review dates: retained evidence and local documentation review through 2026-10-06 UTC.

Review narrative

Role: English STT challenger. Evidence: complete component corpus and paired quality assessment. Decision: no-promotion; a lower point WER is not an established quality winner.

Evidence reviewed: 2026-10-06 UTC. This is a dated snapshot for an English personal, noncommercial, quality-first assistant, not current live-service reporting.

Immutable identity

mlx-community/Qwen3-ASR-1.7B-8bit@a8379a2e2f9e313c9292cdf1af4055ab56d50d55.

Candidate audio runtime: MLX Audio 0.5.8, source 70f4add32911bab6f869b824864ad9f1e24dcb97, MLX 0.32.3. The reviewed Anvil audio launcher file SHA is e7149c09668713835e1cd70725f2af78608ac023c72541f96462c61d7c924661; this file hash is not a Git commit. Runtime distribution/source identity and exact local snapshot gates are retained. License/source screening is recorded separately from local qualification in the source registry.

Tested hardware and topology

Apple M4 Max, 40-core GPU, 48 GiB unified memory. These are direct same-host component HTTP measurements. Private hostnames, addresses, operator identity and paths are redacted. Historical CUDA or other-device evidence is a separate population.

Engine, quantization, KV, context, and concurrency recipe

Community 8-bit conversion, English conditioning, 16-kHz mono WAV. Canonical C1: 24 human cases × 3, with six synthetic cases separate. C4 repeats once; the supplemental 120-human-case holdout repeats once. KV cache and LLM context are not applicable.

Exact reconstruction boundaries · Hardware view

Evidence by measurement class

Status: all corpus requests completed. Measured: canonical human 30/1,077 edits/words = 2.786% micro-WER; warm C1 p50/p95 165.37/314.25 ms (72 requests). Supplemental 58/1,864 = 3.112%. Limits: canonical paired 95% WER-difference interval is −1.770 to +1.084 percentage points, overlapping zero. The critical MISSUS→mister error occurs in all three repetitions. Evidence: STT results and independent assessment.

Decision and promotion state

No replacement is qualified. Keep the existing Parakeet control without newly qualifying it. Its own critical errors do not relax the frozen zero-critical-error gate.

Failures and gotchas

This 1.7B MLX conversion is separate from the historical 0.6B CUDA dossier. Actual microphone/accent, endpointing, physical interruption and combined stability coverage remain missing. The graph uses only the matched human C1 population; the native synthetic/aggregate metrics remain retained.

Public artifacts are declared sanitized derivatives, with original private SHA/bytes in the projection ledger. Completion or documentation does not authorize a service, route, or client-catalog change.

Dated run history