Skip to content

Choose a model

Every TurboQuant build published on Hugging Face, grouped by how much unified memory your Mac has. All are in the TurboQuant-MLX collection.

Speeds are decode speeds as measured on the Mac named. A figure marked * is an estimate for 0.30.0, which made ternary (tqTe) builds decode faster, not a measurement (see Benchmarks). Before downloading, run turboquant-plan --model <repo> to check the fit on your Mac; it reads only the file headers (see Get started).

Reading the Settings column

default means nothing to change. sysctl means raising the GPU memory limit once per boot with sudo sysctl -w iogpu.wired_limit_mb=<value>; the card gives the value. See Memory and macOS settings.

16 GB (Mac mini)

Measured on a 16 GB M4 Mac mini.

Model Build Size Settings Speed on the mini Notes
Qwen3.6-35B-A3B tq3a-tqTe-g64 9.4 GB default ~17 tok/s* Fully in memory. Good for chat; not for agent harnesses.
Qwen3.6-35B-A3B tq3a-tqTe-down4-g64 12.6 GB sysctl 13824 ~17 tok/s* Fully in memory. Passes an Opencode agentic task 3/3 (at 15.2–15.6 tok/s on 0.29.0).
GPT-OSS-20B tq3 9.3 GB default 60–80 tok/s (M-series) ~11 GB peak.
Qwen3.8-27B (vision) tq3-mini-g64 11.55 GiB sysctl 14336 + --prefill-step-size 256 3.7 tok/s 12.95–13.61 GiB peak. Vision passes. Not agentic on 16 GB.
DiffusionGemma-26B-A4B mini-g32 9.79 GB sysctl 13824 ~1.5 tok/s Block-diffusion model. Experimental.

Bigger than 16 GB, by streaming experts from disk (see expert streaming):

Model Build Size on disk Speed on the mini Notes
Qwen3.6-35B-A3B tq3-g32 ~16 GB — Runs in under 4 GB of RAM; output bit-identical to resident.
Qwen3.5-122B-A10B tq3a-tqTe-g64 30.9 GB ~2.1 tok/s 11.3 GB peak, --cache-budget-gb 8.
Qwen3-235B-A22B tq3a-tq2e-g32 70.5 GB ~0.2 tok/s Works, but disk-bound on a mini. Much faster on 64 GB.

24–36 GB

Measured on a 64 GB M4 Max unless noted; these fit a 24–36 GB Mac by size. Check with turboquant-plan.

Model Build Size Speed (M4 Max) Notes
Qwen3.6-35B-A3B tq3-g32 ~16 GB ~60 tok/s ~18 GB peak.
Laguna-XS.2 (33B coding agent) tq3-g64 13.8 GB 46.8 tok/s Fits a 32 GB Mac. Opencode pass.
Qwen3.8-27B (vision) tq4-g64 15.15 GiB 11.3 tok/s Best perplexity of the Qwen3.8-27B builds. Opencode pass, vision 4/4.
Muse Glimmer 30B (vision) tq4-g64 14.88 GiB 9.4 tok/s Better perplexity than affine 4-bit, in 25% less space.
Muse Glimmer 30B (vision) tq3-g64 12.53 GiB 10.9 tok/s Smallest coherent build.
DiffusionGemma-26B-A4B tq3-g32 13.8 GB ~4.6 tok/s ~18 GB peak. Experimental; slower than 4-bit.

48 GB

Model Build Size Speed Notes
Qwen3.6-27B (dense coder) tq3-g32 ~13 GB ~14 tok/s ~17.5 GB at run time; room for long context.
Qwen3.5-122B-A10B tq3a-tqTe-g64 30.9 GB ~25 tok/s* 32.5 GB peak, no sysctl on 48 GB. 6/6 stress.
Laguna-S-2.1 (118B coding agent) tqTe-g64 ~27 GB ~20–28 tok/s* ~30 GB peak. Opencode pass.
GPT-OSS-120B tq3a-tq2e-g32 ~35 GB 42–50 tok/s 3-bit attention, 2-bit experts. 6/6 stress.
Nemotron-3-Super-120B tq3a-tq2e-g32 ~36 GB ~27.2 tok/s ~40.8 GB peak. Math drifts on long generations.

64 GB

Model Build Size Settings Speed (M4 Max) Notes
NVIDIA Nemotron 3.5 Super VL 120B (vision) tq3a-tqTe-g64 29.29 GiB default ~16–21 tok/s Measured on 0.30.0. 44.7 GB peak at a 36K-token prompt. Text 6/6, vision 3/3.
NVIDIA Nemotron 3.5 Super VL 120B (vision) tq3-g64 50.22 GiB sysctl 60416 + --prefill-step-size 256 ~19–21 tok/s About as fast as the ternary build since 0.30.0, but needs both settings.
Qwen3.8-Flash-Next 180B (vision) tq4a-tq2e-g64 52.00 GiB sysctl 60416 16–18 tok/s 56.3 GB peak. --ngram-offload cuts the resident footprint.
Qwen3-235B-A22B tq3a-tqTe-g64 53 GB sysctl 60416 ~9–12 tok/s* Fully in memory, no streaming.
Qwen3.5-122B-A10B tq3 ~54 GB sysctl ~25 tok/s Runs out of GPU memory past ~4–8K context.
GPT-OSS-120B tq3 48 GB check with turboquant-plan 44 tok/s
Nemotron-3-Super-120B tq3 ~50 GB sysctl 18.7 tok/s ~55 GB peak.
Qwen3-235B-A22B tq3a-tq2e-g32 70.5 GB streaming ~4–6 tok/s --cache-budget-gb 40. 5/6 quality probes.
Qwen3-235B-A22B tq3-g32 103 GB streaming ~1.3 tok/s Recall-critical sibling. 6/6 stress.

Which build of the same model?

  • tq3 is 3-bit everywhere: the quality and speed reference.
  • tq3a-tq2e keeps attention at 3-bit and drops experts to 2-bit: smaller, faster, but long generations can drift on some models.
  • tqTe uses ternary (1.58-bit) experts: the smallest build. It needs a model with many wide experts (128+). Since 0.30.0 it decodes close to 3-bit speed (measured on NVIDIA Nemotron 3.5 Super VL); before, it was clearly slower (Qwen3.6-35B: ~33 tok/s ternary against ~60 for 3-bit on an M4 Max).
  • down4 adds a 4-bit expert down projection to a ternary build: a few GB more, and it is what made Qwen3.6-35B work in agent harnesses.
  • g32 / g64 is the group size: 32 is more accurate, 64 is smaller.

Each model page under Models has the details, and each Hugging Face card has the full measurements.