Skip to content

Fish Audio S2 Pro MLX q8

Current status and review date

Decision snapshot

  • Product role: English native-voice TTS 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: streaming incompatibility and resource-rejected buffered full attempt.
  • Decision: no-promotion; keep existing controls without newly qualifying them.
  • Important limitation: 56/60 completion; no partial-cell ranking or human/client qualification.
  • Review dates: retained evidence and local documentation review through 2026-10-06 UTC.

Review narrative

Role: English native-voice TTS challenger. Evidence: unsupported streaming probe and resource-rejected buffered attempt. Decision: no-promotion; partial cells are not ranked.

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/fish-audio-s2-pro-8bit@c8d4481b3f7cbfe64d855c8b7cda7739502fc3ff.

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

Buffered request, default voice unverified, 1,024-token total cap, cache zero and worker trace. Source 44.1-kHz PCM is normalized to 16 kHz. Native chunk_length=300 remains unchanged; no hidden caller chunking. C1, full 60-prompt attempt plus separate cold request.

Exact reconstruction boundaries · Hardware view

Evidence by measurement class

Status: 56/60 warm requests returned; four long-form requests fail at resource abort. Measured: 728,099,718 bytes swap growth exceeds the 536,870,912-byte gate; minimum free memory 28%. Successful-subset ASR micro-WER is 9.443%. Limits: the aborted generation peak is unknown; the 15,286,476,287-byte completed-generation MLX peak does not bound it. Evidence: TTS results, resource assessment and streaming negative probe.

Decision and promotion state

The exact tested configuration is resource-rejected. No human winner or new deployment. A proposed 512-token cap is unmeasured and may truncate; any retry must rerun the unchanged full 60 and all gates.

Failures and gotchas

Pinned upstream streaming raises NotImplementedError; the negative HTTP probe returns zero audio, so it is not a speed comparison. Successful 56-request timing/fidelity cannot qualify the full 60. Cache zero does not cap active allocations; listening and physical client acceptance remain missing.

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