Reference · rapid-mlx 0.10.3 · ← Back to README

Hardware tiers

What to serve based on your Mac's unified memory. These four picks are what install.sh ships as its RAM recommender — one canonical alias per tier, boot-verified against rapid-mlx 0.10.3 on M3 Ultra.

Source of truth. These are the exact aliases the RECOMMENDED_MODEL block in install.sh writes into the quick-start banner after curl -fsSL https://rapidmlx.com/install.sh | bash. Boot times and peak RSS come from PR #1053's cached-model empirical run (M3 Ultra 256 GB, port 8804-8807, /v1/models-Ready gate).

The RAM → alias map

RAM tier Alias (default recommendation) HF repo Boot to /v1/models Peak RSS
96+ GB gpt-oss-120b-mxfp4-q8 mlx-community/gpt-oss-120b-MXFP4-Q8 15 s 60.9 GB
48-95 GB qwen3.6-35b-8bit mlx-community/Qwen3.6-35B-A3B-8bit 11 s 35.6 GB
24-47 GB gpt-oss-20b-mxfp4-q8 mlx-community/gpt-oss-20b-MXFP4-Q8 6 s 12.1 GB
8-23 GB qwen3.5-4b-4bit mlx-community/Qwen3.5-4B-MLX-4bit 8 s 2.8 GB

96+ GB · Mac Studio / Pro

Pick: gpt-oss-120b-mxfp4-q8  ·  rapid-mlx serve gpt-oss-120b-mxfp4-q8

OpenAI's harmony-native 120B MoE at MXFP4 with 8-bit accumulators. Native harmony reasoning + tool calling, so --tool-call-parser harmony is auto-selected. Cache footprint on disk is roughly 59 GB — comfortably below the M3 Ultra 256 GB envelope with headroom for a second sidecar model.

Rationale: at this tier the picky question is "frontier reasoning + tool calling on-device". 120B-MXFP4 is the largest checkpoint that boots in seconds and fits below the 96 GB working-set floor, so it wins the tier over 122B dense / 158B DeepSeek V4 Flash both of which spill past the 96 GB envelope on typical prompts.

48-95 GB · Mac Studio

Pick: qwen3.6-35b-8bit  ·  rapid-mlx serve qwen3.6-35b-8bit

Qwen 3.6 35B-A3B (35B total / 3B active MoE) in 8-bit MLX. Ships with a native MTP head baked into every checkpoint, so --enable-mtp (or a --speculative-config '{"method":"mtp"}') accelerates decode without a separate drafter download. The hermes / qwen3 parser pair is auto-selected.

Rationale: the 35B-A3B sweet spot — active-parameter cost of a 3B model, quality of a 35B — anchors this tier. 8-bit MLX at ~35 GB peak RSS leaves 13-60 GB free at 48-95 GB budgets for prefix cache and concurrent sequences.

24-47 GB · Mac Mini / Studio

Pick: gpt-oss-20b-mxfp4-q8  ·  rapid-mlx serve gpt-oss-20b-mxfp4-q8

The smaller sibling of the 96+ pick. Harmony-native, ~11 GB on disk, 12.1 GB peak RSS during a warm boot. Fits comfortably at 24 GB with a normal browser + IDE workload alongside.

Rationale: at 24-47 GB the ceiling is set by the "leave room for VS Code + Chrome + a build" reality. A ~12 GB peak-RSS harmony-native model that responds at real chat speed on M-series is the honest highest quality that clears this bar.

8-23 GB · MacBook Air / small mini

Pick: qwen3.5-4b-4bit  ·  rapid-mlx serve qwen3.5-4b-4bit

Qwen 3.5 4B at 4-bit MLX. Runs fine on a 16 GB MacBook Air with power to spare. Also the default alias for rapid-mlx chat, rapid-mlx launch, and the matrix cheap-alias policy.

Why not Qwen 3.6 here? Qwen 3.6 has no <8B SKU at time of release, so the recommender falls back to 3.5-4B. When 3.6 ships a small variant, the install.sh map will be re-verified for this tier.

Verify empirically

The installer's doctor subcommand reports your machine's Apple Silicon variant, macOS version, free disk, and HF cache size, then flags any missing optional extras.

$ rapid-mlx doctor

Sample output on M3 Ultra 256 GB:

◆ System
  ✓ Apple Silicon (Apple M3 Ultra, 256 GB)
  ✓ macOS 26.5.1 (Darwin 25.5.0)
  ✓ Free disk: 114 GB
◆ Required Packages
  ✓ mlx 0.32.0
  ✓ mlx-lm 0.31.3
  ✓ rapid-mlx 0.10.3

Next steps