Apple M4 Max benchmark view¶
Hardware: Apple M4 Max laptop, 48 GB unified memory. Reviewed: 2026-09-12, using retained September 8 evidence. Deployment and harness statements are historical, with no current live-state or code-release claim. This page reports only the measured same-host Apple Silicon lane; it does not describe the RTX fleet or the reference Mini-to-Dark topology.
VoiceChat feasibility stop¶
On 2026-09-19, the VoiceChat MLX candidate was inspected but not acquired, loaded, or queried. Its 8.553 GiB package is a disk value only; memory fit is unresolved. The pinned session lacks supported tool-result ingress, blocking a whole tool-capable replacement. Protected audio diagnostics are private-only: no sanitized native artifacts or workload provenance are retained publicly, so they support no public candidate-latency or corpus-quality claim. See the dated finding.
Swift / stock Qwen3.8 artifact feasibility stop¶
The 2026-09-19 read-only screen considered pinned Swift and stock Qwen3.8-27B GGUF Q6_K and Q4_K_M artifacts with their family-matched F16 projectors. With 24.4899 GiB free, a 10 GiB disk reserve, and a 1 GiB runtime/evidence allowance, all four uncached pairs failed the optimistic disk screen before temporary files. The memory calculator was unresolved: peak loader, Metal, temporary-memory, and context requirements were not retained, and no enforced loader/RAM/swap containment exists. No model downloaded or ran; no latency, functional, quality, capacity, vision, or soak evidence exists. The protected voice path reported health 200, but that is not candidate or Talk latency evidence.
| Screen | Result | Decision |
|---|---|---|
| Swift Q6_K + F16 projector | 9,327,321,440 B disk-policy shortfall | stop before download |
| Stock Q6_K + F16 projector | 10,303,479,744 B disk-policy shortfall | stop before download |
| Swift Q4_K_M + F16 projector | 4,467,294,560 B disk-policy shortfall | stop before download |
| Stock Q4_K_M + F16 projector | 3,885,313,984 B disk-policy shortfall | stop before download |
This is a storage-policy and runtime-containment stop, not a claim that either model is impossible on Apple hardware or a cleanup authorization. See the dated finding and its sanitized evidence bundle.
Local voice-lane evidence¶
The 2026-09-08 refresh evaluated an existing Qwen3 4B baseline and MLX 4-bit Qwen candidates against functional, strict spoken, strict controlled-output, and bounded audio/Realtimes gates. No LLM route changed and no LLM candidate was promoted. A separately accepted Kokoro FastAPI 0.8.2 TTS runtime update was deployed. The Qwen3.5-9B candidate passed six functional groups but failed the strict spoken suite 33/36, patch-format 0/3, and controlled output 0/10. Qwen3.8-27B preliminary preflight passed 5/6 groups but shared-prefix tools were 1/3 and the lane was too slow for the voice goal. Qwen3.6-35B-A3B remains an unresolved-policy-memory bounded exploration: its spoken suite passed 36/36 strict and one SDK session passed, but strict JSON preflight failed.
| Capability | Exact configuration | Evidence | Decision |
|---|---|---|---|
| Local voice LLM candidate | mlx-community/Qwen3.5-9B-4bit@8b2b98c00a6b4d291155e4890773ca8f769aee53, MLX-LM 0.31.3 / MLX 0.31.2 |
functional 6/6; diagnostic quality 9/12; spoken 33/36 strict; strict capacity 0/10 |
no-promotion |
| Local voice LLM preliminary lane | mlx-community/Qwen3.8-27B-4bit@3e6447f082e89cc7f0bc6e5441afd38dfce760ff |
compatibility-only preflight 5/6; tools 1/3; no capacity | incomplete, no-promotion |
| Local voice LLM bounded challenger | mlx-community/Qwen3.6-35B-A3B-4bit@38740b847e4cb78f352aba30aa41c76e08e6eb46 |
preflight 5/6 with strict JSON failure; spoken 36/36; one accepted SDK session | incomplete, no-promotion |
| STT/TTS baseline path smoke | Parakeet.cpp TDT 0.6B v3 Metal + Kokoro FastAPI MPS | one synthesized round trip: WER 0.0, 1896.72 ms | path smoke only; not corpus proof |
| Deployed local TTS runtime | Kokoro FastAPI 0.8.2 58b08a915b3463cb76e376a2867e04f9d828f4df; same Kokoro 0.9.4 model/voice assets |
one candidate round trip WER 0.0, 2139.31 ms; final warm Realtime follow-up TTFA/E2E 353.53/2186.88 ms; four endpoints HTTP 200 | deployed bounded runtime update; no LLM promotion or speed ratio |
The Voice LLM MLX dossier separates this Apple lane from Qwen3.8 results on other hardware. See the dated finding and its sanitized evidence bundle.