the new Mac Studio and Mac mini start shipping monday. so this week is one question: which machine do I need for the models people are actually running right now.
how I picked the models: what's being downloaded this month on Hugging Face, what's being talked about on Hacker News and LM Studio, and what has an MLX build. not lifetime downloads, not hype with no adoption.
how to read the sizes: every number is the MLX weights on disk, pulled from the actual file list on Hugging Face. a model needs a bit more than that to run, for the KV cache and macOS. so "fits" means loads with room for normal chat. long prompts on the big models get tight. speed numbers I don't have, that's what the replies are for.
1. the new machines
| machine | memory | bandwidth | from | ships |
|---|---|---|---|---|
| Mac mini M6 | 16 GB base, 32 GB max | 170 GB/s | $899 | sept 22 |
| Mac mini M5 Pro | up to 64 GB | 307 GB/s | $1,699 | sept 22 |
| Mac Studio M5 Max | up to 128 GB | 614 GB/s | $2,499 | sept 22 |
| Mac Studio M5 Ultra | up to 512 GB | 1.2 TB/s | $5,499 | sept 22, 512 GB config late october |
memory decides what you can run. bandwidth decides how fast. pick memory first.
Apple: Mac Studio
Apple: Mac mini
2. the cheat sheet
grouped by the smallest new Mac that fits. within each tier, most downloaded first. size is the MLX weights on disk. dl is Hugging Face downloads in the last 30 days for the original model.
16 GB. base Mac mini M6.
| model | size | why it's here | get it |
|---|---|---|---|
| MiniCPM5 2B (OpenBMB) | 1.4 GB | best model under 4B right now. publisher ships the MLX build themselves | MLX |
| Spark-X2.5 4B | 3.0 GB | 1M context, built for on-device agents | MLX |
| Llama 3.1 8B (Meta) | 4.5 GB | old, still the most pulled model on Ollama by a mile | MLX |
| Gemma 4 E4B (Google) | 5.1 GB | the phone-class Gemma. audio and vision | MLX |
| Gemma 4 12B (Google) | 6.7 GB | 2.7M dl | MLX |
| Ternary Bonsai 2 27B (PrismML) | 8.6 GB | Qwen3.8-27B squeezed to under 9 GB, claims 98% of the quality. two days old, unproven. card reports ~47 tok/s on M5 Max | MLX |
| gpt-oss-20b (OpenAI) | 11.2 GB | a year old, still 6.7M dl. 3.6B active so it's quick | MLX |
32 GB. Mac mini M6 maxed, or M5 Pro.
| model | size | why it's here | get it |
|---|---|---|---|
| Gemma 4 26B-A4B (Google) | 15.3 GB | most downloaded on this whole list, 9.8M dl. 4B active so it flies | MLX · Unsloth 3-bit, 13.2 GB |
| Qwen3.8-27B (Alibaba) | 16.1 GB | 7.4M dl, top of the agentic leaderboards, does vision | MLX |
| Qwen3-Coder-30B-A3B (Alibaba) | 17.2 GB | the coding workhorse. 3B active | MLX |
| Gemma 4 31B (Google) | 18.4 GB | the dense one | MLX · Unsloth 4-bit, 23.3 GB |
| Qwen3.6-35B-A3B (Alibaba) | 20.4 GB | or 14.9 GB pruned, see section 3 | MLX · pruned |
| MiniMax H3 (video) | 25.3 GB | Hailuo 3.0 video model, open since august, 4.4M dl. 15 s clips with stereo audio. community conversion. check the license, reportedly restricts public use in US, EU, UK | MLX |
128 GB. Mac Studio M5 Max.
| model | size | why it's here | get it |
|---|---|---|---|
| gpt-oss-120b (OpenAI) | 62.3 GB | 5.2M dl. this alone justifies a Max | MLX |
| MiniMax M2 pruned (139B) | 73.9 GB | see section 3 | MLX |
| GLM 4.7 pruned (185B) | 98.2 GB | see section 3 | MLX |
| Tencent Hy3 (295B) | 98.8 GB | 2.7 bits, Apache 2. a 295B model inside a Max | MLX |
256 GB. Mac Studio M5 Ultra.
| model | size | why it's here | get it |
|---|---|---|---|
| DeepSeek V4 Flash | 160.1 GB | last week's V4.1 Flash only has a 428 GB build so far, that one needs the 512 | MLX |
| Kimi K3 (Moonshot) | 181.2 GB | pruned to 543B, 2-bit. the biggest open model ever released. see section 3 | MLX |
| GLM 5.3 Flash (Z.ai) | 181.9 GB | mixed 4/8-bit. 320B total, 18B active, 2.7M dl | MLX |
| MiniMax M3 (428B) | 241.5 GB | official mlx-community build | MLX |
512 GB. late october.
| model | size | why it's here | get it |
|---|---|---|---|
| GLM 5.3 (Z.ai) | 419.4 GB | the strongest open coding model on the benchmarks | MLX |
| DeepSeek V4.1 Flash | 427.6 GB | mixed build. expect a smaller one soon | MLX |
short version: a 32 GB Mac mini runs the five most downloaded models in the world right now. a Max adds gpt-oss-120b and the pruned giants. a 256 Ultra runs everything but the two 400 GB monsters.
3. what you couldn't run last month and can now
some of those sizes only exist because of a pruning trick that showed up on mlx-community in the last few weeks.
REAP, from Cerebras, ICLR 2026. a MoE model has hundreds of experts per layer and most barely get used. REAP ranks each expert by how often the router picks it times how much it changes the output, then deletes the bottom half. then you quantize what's left. the paper claims 96.8% of code quality kept at 25% pruning and 91.4% at 50%, on their test models. nobody has benchmarked the community builds below yet, so treat those as the paper's claim, not a promise.
what it changed:
Qwen3.6-35B-A3B: 20.4 GB down to 14.9 GB, with the card reporting 11.6 GB peak. tight to comfortable on a 24 GB Mac.
Kimi K3: 2.78 trillion parameters, no MLX build possible. pruned and 2-bit it's 181 GB on a 256 Ultra.
MiniMax M2 and GLM 4.7: both were Ultra-only. pruned they're 74 and 98 GB, inside a Max.
there's a small tool that does this to any mlx-lm MoE on your own Mac.
reap-mlx on GitHub
4. OptiQ, the other name in those repos
OptiQ is a mixed precision quantizer for MLX. sensitive layers get more bits instead of everything at 4. pip install mlx-optiq.
the part that matters on a small Mac is the KV cache. their post claims a 4-bit KV cache at 32k context runs 34% below fp16 and within 2% of its speed, while stock mlx-lm's 4-bit cache actually uses more than fp16 at that length.
if you're on 24 or 32 GB and running long context, try this first.
OptiQ docs
the KV cache post
5. Unsloth now ships MLX weights
if you grab Unsloth GGUFs out of habit, they now publish native MLX builds, tagged UD for Unsloth Dynamic. Qwen3.6-35B-A3B, Qwen3.6-27B, Gemma 4 31B and 26B-A4B so far. linked above where they exist.
Unsloth on Hugging Face
6. video on a Mac, quick status
MiniMax H3 is the one to try, see the 32 GB tier. Wan 2.2 from last year still runs through mlx-video, but Wan 3.0 launched in august as closed API only after Alibaba pre-announced it as open. Seedance 2.5 is API only. HunyuanVideo 1.5 has an unofficial Mac port and not much else.
mlx-video
7. also this week
mlx-vlm shipped 0.7.0 and 0.7.1: continuous batching, MoE offloading, GLM 5 Next, SAM 3.1 video tracking, DINOv2.
mlx-audio shipped three point releases with streaming speech to text and new TTS models.
mlx-lm has no new tag yet but the last week of commits adds OLMo Hybrid, Spark-X2.5, a Qwen3 Next pipeline, a server memory leak fix on model swap, and a faster detokenizer. expect a release soon.
mlx-vlm releases
mlx-audio releases
mlx-lm commits
that's the week.
if you ordered one of the new machines, which tier did you pick and what's the first model you're loading?
and if you're already running any of these on Apple silicon, post your chip, the model, and tok/sec. I'll fold real numbers back into this post.
Stay Frosty,