Edge0 35B-A3B on a 16GB Mac mini (2026): Still Buy 24GB

Bottom Line

• A 16GB Mac mini can probably load Edge0-35B-A3B. The model streams experts from SSD and reports 2.9 GiB peak active memory.
• The speed claim was not measured on 16GB. Edge0 measured 14.9-17.7 tok/s decode on a Mac mini M4 Pro with 24GB. No 16GB result is published.
• That difference matters. The checkpoint is 18.4 GiB. On 24GB, most of it can stay in the macOS file cache after the first request. On 16GB, it cannot, so more reads go to the SSD.
• You pay 3.9 benchmark points for the trick. Edge0 scores 79.2 on average against 83.2 for the fp16 base, by the vendor’s own tests.
• Our advice: if you already own a 16GB Mac, try it for free. If you are buying, buy 24GB or more. RAM still decides what runs well.

What Edge0-35B-A3B Is

All figures come from the model card, the GitHub README and the Hugging Face API, read on 18 September 2026.

The README lists the checkpoint as “~23 GB”. The Hugging Face file list sums to 19.5 GB for the shards plus 0.18 GB of adapters. Plan for 25 GB of free disk either way.

How SSD Expert Streaming Works

A normal runtime loads all 35B weights into memory. Only about 3B are used per token, but all of them must be ready.

Edge0 keeps the expert weights on disk and maps them in on demand. The streaming docs say each layer’s MoE weights are about 310 MB at 4-bit. A small cache holds recently used experts.

The hard part is timing. The arXiv abstract explains that the next layer’s experts are not known until the current layer finishes. So a plain SSD read always stalls. Edge0 adds a prerouter that predicts the next layer’s experts one token ahead. The paper says the prediction “is consumed as the routing itself”. In other words, the predicted experts are the ones that run.

That design is why a quality cost exists. The LoRA adapter is trained to pay back the loss from int4 and from the replaced routing. The README says the prerouter adds up to +59% decode speed.

The Number Nobody Else Is Pointing At

The README defines its two prefill numbers this way: cold is the first request, when expert weights “fault in from SSD”. Warm is later requests, with the “page cache resident”.

Page cache is RAM. macOS keeps recently read file data in free memory and does not count it as the app’s memory. So “2.9 GiB active” is true, and the machine can still be using many more GB of RAM to make it fast.

On the 24GB test Mac, an 18.4 GiB checkpoint can mostly sit in that cache. On a 16GB Mac, it cannot. The OS, your browser and the KV cache need room too. So a 16GB Mac will read more from SSD per token than the benchmark machine did. We expect it to be slower. We have not measured how much slower, and Edge0 has not published it.

Two more gaps between the headline and the fine print:

• The speed figures do not agree. The model card and README say 14.9-17.7 tok/s. The arXiv abstract says 20 tok/s “on a single 24GB machine”.
• The test chip is an M4 Pro. The README lists supported chips as M1 to M4. It does not mention the M5 Pro or M6 Mac minis Apple launched on 25 August 2026.

Edge0 vs Plain Qwen3.6-35B-A3B in RAM

The alternative is the normal GGUF, fully in memory. Sizes come from the Unsloth file list, read on 18 September 2026.

macOS gives the GPU only part of unified memory by default, so “tight” means short context. On 16GB, the plain GGUF route forces you down to 2-bit. That is where Edge0 has a real case: it gives a 16GB Mac a 4-bit 35B model with a trained repair adapter.

We did not find a dated, named Mac speed report for plain Qwen3.6-35B-A3B that we could cite. Measure it yourself with the method below.

Honest Tradeoffs

1. It is a preview. The model card says tool use, multi-step planning and long-horizon agent work “are currently weak”. That rules it out as an OpenClaw agent model for now.
2. Quality drops 3.9 points on average. The widest gap is AIME 2026: 86.6 vs 92.7. HumanEval drops from 95.1 to 90.9.
3. No Ollama, no llama.cpp. It runs only through edge0 on MLX. It pins mlx==0.30.6 and mlx-lm==0.31.0.
4. Long context costs RAM. The card says long contexts grow the KV cache. The 2.9 GiB figure is for short contexts.
5. SSD speed becomes your model speed. The README says the prerouter gain “grows with storage latency”. A slow external drive will hurt. Keep the model on the internal SSD.
6. SSD wear is mostly a non-issue. This workload reads the SSD. Flash wear comes mainly from writes. The real write risk on a 16GB Mac is swap, if other apps push memory over the limit.

How to Test It on Your Own Mac

This costs nothing but disk space and an hour.

1. Install it: pip install -e 'git+https://github.com/Edge0-AI/edge0.git#egg=edge0[fetch]'
2. Download: huggingface-cli download Edge0/Edge0-35B-A3B-preview --local-dir ./edge0-35b
3. Run the vendor benchmark: EDGE0_35B_MODEL=$PWD/edge0-35b python examples/bench.py edge0-35b from the cloned repo.
4. Open Activity Monitor. Watch Memory Pressure and Swap Used, and the Disk tab’s Data read/sec.
5. Run it twice. The first run is cold. The second shows how much the file cache helps on your RAM size.
6. Compare the result against 14.9-17.7 tok/s. If your decode is far lower and memory pressure is yellow or red, RAM is your limit, not the model.

What to Buy

If you already own a 16GB Apple Silicon Mac, try Edge0. It is the only way we know to run a 4-bit 35B model on that machine.

If you are buying for local LLMs, do not plan around streaming. Buy RAM. Prices below are as of September 2026.

Apple does not publish per-step memory prices in a form we could verify. Check the configurator for the 24GB and 32GB options.

The cheapest way into Apple Silicon is the Mac mini M4, while retail stock lasts. Pick the 24GB version if you can find it. It caps at 24GB, and it is the base M4, not the M4 Pro that Edge0 tested.

For the full tier question, see the local LLM estimator or the gear page.

Sources

• Edge0/Edge0-35B-A3B-preview model card — specs, quality table, benchmark machine, limitations (read 18 September 2026)
• Hugging Face API file tree for Edge0-35B-A3B-preview — shard and adapter sizes (read 18 September 2026)
• Edge0-AI/edge0 GitHub README — supported OS and chips, cold/warm definitions, benchmark method (read 18 September 2026)
• edge0 streaming docs — per-layer expert size, cache design (read 18 September 2026)
• arXiv 2609.18063 — abstract, 20 tok/s on a 24GB machine, submitted 16 September 2026
• Hugging Face API file tree for unsloth/Qwen3.6-35B-A3B-GGUF — plain GGUF sizes (read 18 September 2026)

See Also

• Best Local LLM for Mac mini — why 24GB is the floor for local models on a mini
• Mac mini vs GPU for Local LLM — the same money spent on a graphics card
• Mac mini M4 Pro: 48GB vs 64GB — the higher memory tiers
• llama.cpp MoE Offload Flags Explained — the GPU-side version of expert offload
• Best 20B to 35B Local LLMs — where Qwen3.6-35B-A3B sits among its peers
• Best Local LLM for a 16GB MacBook Pro — the 16GB problem on a laptop