M4 Max local voice refresh: functional candidates, retained strict failures

Date: 2026-09-08

Publication review: 2026-09-12. All service, deployment, and harness statements below describe September 8 observations. This publication reuses retained sanitized artifacts; it makes no live request, deploys no code, and does not attest to current fleet state. The capture used dirty source, and its historical evidence-author model is unrecorded. The earlier native-dispatch and SDK-harness edits are outside this documentation-only change; their current release status is not established by these artifacts.

Measured hardware: Apple M4 Max laptop, 48 GB unified memory

Topology: same-host local voice lane. STT was Parakeet.cpp TDT 0.6B v3 on Metal and TTS was Kokoro FastAPI on MPS. This does not measure the reference Mini-to-Dark proxy topology.

Decision: no-promotion for the LLM lane. The existing Qwen3 4B LLM configuration remains unchanged. The bounded candidate sweep is complete; a separately verified Kokoro TTS runtime update was deployed.

Result card

On one Apple M4 Max laptop, pinned Qwen3.5-9B passed functional preflight but failed strict spoken, patch-format, and controlled-output gates; it is not a performance-qualified voice replacement.

Setup	Local result
Baseline	mlx-community/Qwen3-4B-Instruct-2507-4bit@50d427756c6b1b2fe0c0a10f67fbda1fc8e82c1b; MLX-LM 0.31.3 / MLX 0.31.2
Candidates	Qwen3.5-9B 8b2b98c00a6b4d291155e4890773ca8f769aee53; Qwen3.8-27B 3e6447f082e89cc7f0bc6e5441afd38dfce760ff; Qwen3.6-35B-A3B 38740b847e4cb78f352aba30aa41c76e08e6eb46
Measurement path	direct loopback endpoints with retained mixed cold/warm preflights; 4K/C1 strict unique-cache capacity scout; one synthesized STT/TTS round trip
Contract	12 spoken-assistant cases × 3 repeats; 32-word controlled output, 10 requests; thinking disabled
Evidence	local functional, diagnostic quality, failed capacity; Qwen3.8 compatibility-only evidence is incomplete
Decision	no-promotion for LLMs; Qwen3 4B LLM configuration unchanged; Kokoro 0.8.2 TTS runtime deployed separately

Headline measurement	Local result	Conditions
Functional preflight	6/6 pass	Qwen3.5-9B: smoke, JSON, 4K needle, tools 3/3, streaming tools, continuation
Spoken suite	33/36 strict	Qwen3.5-9B; Tea. failed the native exact tea validator
Patch-format diagnostic	0/3	Qwen3.5-9B built-in quality subset
Strict capacity	0/10	4K/C1, 32 requested words; canaries 10/10, first retained sample observed 28 code words
Qwen3.6 spoken suite	36/36 strict	bounded exact-question suite; separate strict JSON preflight failure remains
Synthesized audio round trip	WER 0.0; 1896.72 ms	one Parakeet/Kokoro sample only
Kokoro 0.8.2 candidate round trip	WER 0.0; 2139.31 ms	one synthesized sample only; candidate endpoint, not a speed comparison
Deployed TTS Realtime follow-up	TTFA 353.53 ms; E2E 2186.88 ms	one warm session after repeated captures; no speed ratio

Why it matters: functional compatibility alone would conceal strict output failures that make a voice-lane performance comparison invalid.

Important caveat: the audio measurement is one synthesized sample, not a corpus or perceptual-quality proof. Qwen3.8 and Qwen3.6 exploration do not change this point-in-time decision.

Evidence: artifact manifest · evidence index · publication summary

Immutable identity and scope

The retained 4B baseline is mlx-community/Qwen3-4B-Instruct-2507-4bit@50d427756c6b1b2fe0c0a10f67fbda1fc8e82c1b on MLX-LM 0.31.3 and MLX 0.31.2. The completed candidate is mlx-community/Qwen3.5-9B-4bit@8b2b98c00a6b4d291155e4890773ca8f769aee53. Artifact verification passed for all three candidate conversions. Thinking was disabled with enable_thinking=false through the chat template; this is an observed control, not exhaustive runtime attestation.

Qwen3.8-27B uses mlx-community/Qwen3.8-27B-4bit@3e6447f082e89cc7f0bc6e5441afd38dfce760ff. It is a separate Apple Silicon lane from the existing large Qwen3.8 dossier, which records other hardware. Qwen3.6-35B-A3B uses mlx-community/Qwen3.6-35B-A3B-4bit@38740b847e4cb78f352aba30aa41c76e08e6eb46; the Hub metadata was created or modified 2026-04-16. Its native qwen3_5_moe MLX architecture is supported; the managed private service manifest declares a 24,576 MiB budget, while the conservative 24 GiB policy reserve interval overlaps demand. It is bounded exploration only, not a measured-memory or all-gates-pass claim.

Functional gates

The existing 4B baseline failed the shared-prefix tools group 0/3, while its spoken suite passed 36/36 strict. Qwen3.5-9B passed smoke, JSON, a 4K needle, shared-prefix tools 3/3, streaming tools, and tool-result continuation. These six functional gates establish compatibility, not a promotion decision.

Qwen3.8 preliminary preflight passed 5/6 groups. The shared-prefix tool group passed 1/3; its one retained successful sample took 35.463 seconds at 5,481 prompt tokens. The two failed exceptions were not individually serialized, so this finding does not assign a failure cause. It was too slow for the local voice goal and did not advance to capacity qualification.

Qwen3.6-35B-A3B preliminary preflight also passed 5/6 groups. It passed tools 3/3, streaming tools in 0.7 seconds, tool-result continuation in 0.8 seconds, and a 4K-target needle in 2.5 seconds, but failed strict JSON: the model's answer was {"error":"Invalid request. The user's input does not match the required format."} rather than the required language/ok object. This is a bounded format failure, not a universal model-quality conclusion. Spoken and SDK voice results then completed: the spoken suite passed 36/36 strict, and one SDK session had no acceptance errors, successful cancellation, and completed follow-up 54.. The cancelled response TTFA/latency was 3890.11/3907.36 ms; the follow-up was 484.69/664.81 ms. This is N=1 with different warm/cache histories, so it does not support a cross-model speed ratio. The strict JSON failure persists and blocks qualification.

Negative results retained

The Qwen3.5-9B diagnostic built-in quality score was 9/12, with patch-format 0/3. Its spoken assistant suite was 33/36 strict because all three memory-case answers were Tea. instead of the exact expected tea. That answer is semantically correct, but the native evidence remains strict and was not regraded. This tiny exact-question suite does not establish broad inferiority to the baseline's 36/36 result.

The 4K/C1 unique-cache capacity scout requested 32 controlled output words in ten requests. All ten canaries passed. The first retained sample observed 28 code words, and all ten requests failed strict adherence. The retained evidence does not support calling this an OOM or parser failure. It makes every request performance-ineligible, so no throughput or latency headline is published.

Voice path and Realtime protocol

The Parakeet/Kokoro synthesized round trip returned WER 0.0 in 1896.72 ms for one sample. It is useful as an end-to-end path smoke only; it is not a corpus accuracy, latency distribution, or speech-quality claim.

The retained STT runtime is Parakeet.cpp tag v0.5.0 at 1bfbebfaaf493866f49597cd3b7901959d395c60, with local ggml-submodule modifications retained in the native runtime record. The campaign started with Kokoro FastAPI 0.8.0rc1 at 577595854864fa014b041f5120a84810558942dc. Upstream v0.8.2, released 2026-09-05 at 58b08a915b3463cb76e376a2867e04f9d828f4df, was evaluated in an isolated frozen-lock environment and then deployed to the baseline TTS port. Its one synthesized candidate round trip had WER 0.0 in 2139.31 ms (STT 94.66 ms, TTS 2044.66 ms); this is one path smoke, not a quality or speed comparison. The initial candidate definition failed warmup because relative asset paths resolved from the repository root; the retained native failure records the error and the pinned definition was corrected with absolute paths. The deployed runtime keeps the Kokoro 0.9.4 model and voice asset hashes unchanged. Versioned TTS/proxy supervisor definitions use KeepAlive=true and ThrottleInterval=10; prior definitions were disabled and retained for rollback.

The Qwen3.5-9B official OpenAI Python SDK capture initially observed one inflight transcript after response.cancel was sent. The historical dirty-worktree harness treated the standard server response.done event with cancelled status as the cancellation fence. The current main-branch harness still uses the client-send fence; this publication does not change or qualify that behavior. Subsequent September 8 baseline and Qwen3.6 captures passed the modified fence with their own model configurations; they are not a same-model Qwen3.5 repeat. The original Qwen3.5 negative capture remains retained as diagnostic evidence.

The existing long-running managed proxy then passed the corrected fence with no acceptance errors: both audio utterances were recognized, no event appeared after either the client cancel request or the cancellation acknowledgement, and the completed answer 54. had TTFA 1678.46 ms and end-to-end latency 2740.76 ms. The first cancelled response TTFA was 3559.19 ms. Exact package versions and source-file hashes remain in the private runtime receipt. The acknowledgement boundary follows the official Realtime cancel event documentation.

After the TTS deployment, a fresh corrected SDK wrapper exited zero with no acceptance errors, observed the cancelled acknowledgement, and completed the follow-up answer 54 countries in Africa. at TTFA 353.53 ms and end-to-end 2186.88 ms. It is one warm session after repeated captures, so it provides no cross-run speed ratio. Two earlier wrapper invocations raised Python invocation errors after valid captures; they are retained privately and are not described as pipeline failures. Reboot, physical microphone/OpenClaw client, and actual rollback execution were not tested.

Feasibility and research

The local feasibility screen uses an 8,192-token requirement and a 24 GiB host/co-resident reserve on 48 GB unified memory. Qwen3.5-9B and Qwen3.8-27B were benchmark survivors; Qwen3.6-35B-A3B crossed the conservative policy interval and remains unresolved despite a physical-fit bound. This screen is not behavioral proof.

Official Qwen repositories, the exact MLX community conversion repositories, LiquidAI LFM2.5-8B-A1B, and Hugging Face speech-to-speech 1.0.0 were reviewed on 2026-09-08. The LFM item is only a future Pythonic-tools/reasoning tradeoff lead. Hugging Face speech-to-speech 1.0.0 was released 2026-09-06, but Anvil uses its native pipeline, so no blind package update occurred.

Decision and promotion boundary

No model is promoted. All three candidates were unloaded and unregistered; both candidate ports are closed. The 9B candidate failed strict gates, Qwen3.8 was too slow with an incomplete tool group, and Qwen3.6 retained its strict JSON failure. Qwen3.6 is a stronger bounded spoken challenger than 9B, but not a fully qualified replacement because JSON failed. The existing 4B LLM configuration has not changed. Four local jobs were adopted in the private service inventory; the stopped proxy was repaired with ordered dependencies and STT /health readiness. Kokoro FastAPI 0.8.2 and versioned TTS/proxy supervisor definitions are the accepted audio-runtime changes. Baseline LLM, STT, TTS, and Realtime endpoint checks all returned HTTP 200. The live voice configuration bytes are unchanged since the audio update began; this does not claim identity with an unavailable pre-campaign byte digest.

Evidence and redaction

The sanitized bundle keeps preserved native JSON plus derived navigation summaries, exact revisions, metrics, and material failures. It replaces personal paths with <operator-home> or <model-cache>, host identity with apple-m4-max-laptop, and excludes audio, base64, and raw SDK streams. Private raw evidence remains operator state.