mlx-dspark

Models

Supported models

Every model here is measured and vouched for. Pass it to --model and its drafter, and that pair's best settings, resolve automatically, for any quantization of the model.

These are the pairs we have measured. They're also exactly what --model auto-resolves, so this list and the engine can't disagree: the page is generated from the engine's registry. Numbers are an M4 Pro's. Your Mac derives its own settings on first run.

Show
ModelMean speedupChat / code / mathTokens / sMemoryDefault drafter
Qwen3.8-27B 8-bit 3.74× 3.07 / 3.93 / 4.23 25–35 ~29 GB DFlash 2
LFM2.5-1.2B bf16 3.35× 2.40 / 3.83 / 3.82 239–381 ~4 GB DSpark
Gemma-4 12B 8-bit 3.25× 2.62 / 2.90 / 4.24 48–78 ~15 GB DSpark
MiniCPM5-2B bf16 3.16× 2.12 / 3.15 / 4.21 107–213 ~6 GB DSpark
Nanbeige4.2-3B bf16 2.92× 1.98 / 2.50 / 4.31 34–75 ~10 GB DSpark
LFM2.5-2.6B bf16 2.79× 2.24 / 2.65 / 3.49 98–152 ~7 GB DSpark
Qwen3.8-27B 4-bit 2.72× 2.05 / 3.12 / 3.02 30–46 ~18 GB DFlash 2
Muse-Glimmer-30B 8-bit 2.47× 1.97 / 2.45 / 2.99 16–25 ~40 GB DSpark (community)
Ornith-1.0-9B 8-bit 2.40× 2.21 / 2.53 / 2.48 59–68 ~13 GB DSpark (community)
Qwen3.6-27B 8-bit 2.29× 2.26 / 1.96 / 2.67 16–22 ~32 GB DSpark (community)
Qwen3-8B 8-bit 2.19× 1.71 / 2.03 / 2.83 50–83 ~11 GB DSpark
Qwen3-14B 8-bit 2.03× 1.62 / 2.11 / 2.36 25–36 ~19 GB DSpark
Qwen3-4B 8-bit 1.98× 1.78 / 1.72 / 2.46 89–127 ~8 GB DSpark
Muse-Glimmer-30B 4-bit 1.74× 1.57 / 1.70 / 1.94 22–27 ~26 GB DSpark (community)
Qwen3.6-35B-A3B 4-bit 1.32× 1.05 / 1.24 / 1.67 91–145 ~23 GB DSpark (community)
LFM2.5-8B-A1B bf16 1.26× 1.04 / 1.30 / 1.44 68–94 ~19 GB DSpark
Nemotron-3.5-Lightning-30B-A3B 4-bit 1.10× 0.95 / 1.23 / 1.13 87–112 ~20 GB DSpark (NVIDIA)
Ternary-Bonsai-27B 2-bit (ternary) 1.07× 1.01 / 1.13 / 1.07 26–29 ~12 GB DSpark

Mean speedup is over plain decoding of the same model, averaged across a chat, a code and a math prompt. Tokens / s is the range across those prompts: chat is usually the low end, because open-ended text accepts fewer drafted tokens than code or math. Memory is the measured peak at chat length for the model plus its 4-bit drafter; long contexts add KV cache on top. How these are measured.

Run one#

mlx-dspark serve --model mlx-community/Qwen3.8-27B-4bit

Matching ignores quantization: a -4bit, -8bit or -bf16 build of the same model resolves the same drafter. mlx-dspark models prints this list in your terminal.

Anything else runs too#

This list is what we vouch for, not what works:

Where models come from#

--model takes a Hugging Face repo id or a local path, like mlx-lm. Before downloading anything, the engine looks, in order, at:

  1. the path itself, if it's a folder (--model ~/models/Qwen3.8-27B-4bit always works);
  2. ~/.cache/mlx_dspark/models;
  3. LM Studio's folder (~/.lmstudio/models): its MLX downloads load in place, with no second download;
  4. any folders you list in MLX_DSPARK_MODEL_DIRS (:-separated), such as an external drive, ~/models or a NAS, as publisher/model trees, publisher_model folders or bare model folders;
  5. your Hugging Face cache (wherever HF_HOME puts it), and only then a download.

Only MLX checkpoints load from these folders (a config.json plus .safetensors). mlx-dspark doctor --json lists every folder searched, in order.

Downloads are big

A 27B model at 8-bit is about 29 GB. Each model page lists what you're signing up for, and the Mac app shows download size and fit before it starts.

Suggest an edit to this page