SPB Git

spb/zyquo-local Public MIT

Native macOS AI chat that runs LLMs 100% locally on Apple Silicon with MLX — no cloud, no API keys.

macosswiftaimlxlocal-llm
Swift 97.2% Shell 1.8% Makefile 1%
2.0 KB
View source Raw Blame History

# Phase 7 — Local Model Verification

Run on 2026-07-30 with ZyquoLocal --verify on the build machine (M5 Max, 48 GB, macOS 27.0 beta). Per model: download through the app's own DownloadManager, directory validation, load, deterministic generation (temperature 0, "Reply with exactly: OK" → replied "OK"), multi-turn context carry-over ("vermilion" recall), streaming-cancellation test (stream stops in ≤0.4 s after cancel), unload with verified memory release (MLX active memory drops to KBs).

# End-to-end results — 6 models across architectures

Model Download Load Generate Multi-turn Cancel Unload tok/s TTFT
Qwen3-0.6B-4bit ✅ ✅ ✅ ✅ ✅ ✅ 109.4 0.08s
Llama-3.2-1B-Instruct-4bit ✅ ✅ ✅ ✅ ✅ ✅ 222.8 0.74s
SmolLM3-3B-4bit ✅ ✅ ✅ ✅ ✅ ✅ 97.4 0.11s
gemma-3-4b-it-qat-4bit ✅ ✅ ✅ ✅ ✅ ✅ 58.3 1.08s
Qwen2.5-Coder-7B-Instruct-4bit ✅ ✅ ✅ ✅ ✅ ✅ 112.5 0.42s
DeepSeek-R1-0528-Qwen3-8B-4bit ✅ ✅ ✅ ✅ ✅ ✅ 61.3 0.08s

Architecture coverage: qwen3, llama, smollm3, gemma3, qwen2 (coder), R1 distill (reasoning, <think> streaming). Memory release examples: 4.65 GB → 20 KB (R1-8B), 2.63 GB → 13 KB (gemma-3-4b).

Per the protocol, test models except the smallest (Qwen3-0.6B-4bit, kept for ongoing dev) were deleted afterward to reclaim disk.

# Featured catalog dry-verification — 30/30 ✅

Every repo in docs/MODELS.md / ModelCatalog verified against the live Hub: repo exists, required files listed, total size within 10 % of the catalog value (all matched to the hundredth of a GB; no gated repos, no missing files, no corrections required — MODELS.md and ModelCatalog remain in sync).

VERDICT: ALL GREEN (--verify exit 0)