Qwen3.8 35B-A3B on a Mac mini: How Much RAM Do You Actually Need?

Aug 31, 2026 · Qwen · Mac mini · Local LLM · 5-min read | All numbers vendor-reported or author-estimated, labeled as such

I almost skipped the Qwen3.8 launch to chase the 27B dense hype. Then a quiet commit in Alibaba's ms-swift repo popped up: Qwen/Qwen3.8-35B-A3B — 35B total, only 3B active. Same MoE recipe that made the 3.6-35B-A3B a darling for 8–16GB VRAM rigs, now on the 3.8 architecture. Not officially announced yet, sitting in the supported-models table. If you're shopping for a Mac mini to run it locally, this is the model you actually want.

The short answer: Q4_K_M (~17.5GB) fits a 24GB Mac mini M6 with room for 32K context and does ~56 tok/s. It does not fit a 16GB mini cleanly. If you want Q8 quality (~35GB), you need a 64GB Mac mini M5 Pro — that's where Q8 unlocks ~51 tok/s (and Q4 flies at ~102 tok/s). 32GB M6 is the comfortable headroom pick on the M6 line.

I'm writing this as a buyer, not the Qwen team. Prices are Apple US 2026 (M6 $899, M5 Pro from $1,699); KR street pricing moves with the DRAM shortage — check Danawa before you buy. Performance numbers are estimates from bandwidth ÷ active-params with a 0.4–0.6 real-world factor.

Wait — 35B-A3B? Isn't Qwen3.8 a 27B?

Qwen's naming is total-Aactive for MoE, plain number for dense. So:

Model	Type	Total / Active	Best for
Qwen3.8-27B	Dense	27B / 27B (reads all)	Best raw quality at this size, but slow — ~11 tok/s on M5 Pro Q4
Qwen3.8-35B-A3B	MoE	35B / ~3B (reads 3B/token)	Same Q4 RAM (~17.5GB) but ~9× faster — the local-agent sweet spot
Qwen3.8-Flash-Next	MoE	125B / 6B	Needs 128GB+ unified — different tier, not this post

There is no "27B-A3B." If someone offers you a 27B MoE, they mean the 35B-A3B — same 3B active (identical speed), more total params (better quality), near-identical Q4 memory. On a Mac mini, the A3B is the point.

Why MoE matters on a Mac: memory = total params, speed = active params. A 35B MoE must fit 35B in RAM (like a 35B dense), but each token only reads 3B — so it feels like a 3B model at decode. That's the whole trick.

How Big Is It Really? (By Quant)

The native BF16 is ~70GB (35B × 2 bytes) — forget it on a mini. The practical rungs come from Unsloth/Greets on the 3.6-35B-A3B precursor — the 3.8-A3B will land within a few %:

Qwen3.8 35B-A3B quant sizes vs Mac mini RAM

Quant	Weights	Total at 32K ctx*	Fits on
Q4_K_M	~17.5GB	~19.5GB	✅ 24GB / 32GB / 64GB — sweet spot
Q5_K_M	~21.5GB	~23.5GB	✅ 24GB tight, 32GB comfy
Q6_K	~26GB	~28GB	✅ 32GB, ✅ 64GB
Q8_0	~35GB	~37GB	❌ 16/24/32GB — ✅ 64GB only
BF16	~70GB	~72GB	❌ minis — needs Mac Studio

* +KV cache + ~1GB runtime overhead. At 32K, add ~2GB; at 128K add ~7.5GB (see Fig.3). Long 262K context on Q4 needs 32GB minimum.

What 32K / 128K Context Really Costs

People budget weights and forget KV. For a 35B-A3B Q4 build, the KV grows with context. At 8K it's noise; at 262K it's a second model:

How Fast on Each Mac mini?

Decode on Apple Silicon is bandwidth-bound: tok/s ≈ bandwidth ÷ bytes-read-per-token with a 0.4–0.6 real-world factor (MoE routing, attention). For Q4, bytes/token = 3B × 0.5 = 1.5GB.

Config	BW	Math (theory)	Realistic	Feels like
A3B Q4 on M6	170GB/s	170 ÷ 1.5 = 113 tok/s	~45–65 tok/s	Smooth chat, fast agent
A3B Q4 on M5 Pro	307GB/s	307 ÷ 1.5 = 204 tok/s	~80–120 tok/s	Very fast — best value speed
A3B Q8 on M5 Pro	307GB/s	307 ÷ 3.0 = 102 tok/s	~40–60 tok/s	Slower than Q4, better quality
27B dense Q4 on M5 Pro	307GB/s	307 ÷ 13.5 = 22.7 tok/s	~9–14 tok/s	Noticeably laggy vs A3B

The punchline: A3B Q4 and 27B Q4 need the same* RAM (~17GB) — but A3B is ~9× faster because it only reads 3B active. If your workload is agentic coding/chat on a mini, speed wins.

A3B MoE vs 27B Dense — Same RAM, Different Machine

Qwen's own 3.6 numbers tell the tier story: the dense 27B scores a bit higher on pure bench (SWE-bench Verified 77.2 vs A3B's 73.4) but the A3B's speed makes it the better real agentic coder on a mini — the dense can't keep up in a tool-loop where you pay per-token latency.

So Which Mac mini Should You Buy?

Mini	RAM / BW	A3B Q4 Verdict	A3B Q8 Verdict	Who it's for
M6 16GB ($899)	16GB / 170GB/s	❌ Tight — 17.5 > 16, needs offload/swap, slow	❌ No	Skip for A3B — general use mini, not local LLM
M6 24GB	24GB / 170GB/s	✅ Sweet spot — comfy at 32K, ~56 tok/s	❌ No	Most buyers — best perf/$ for A3B
M6 32GB	32GB / 170GB/s	✅ Headroom — 64K–128K comfy	❌ No	Long-context users who stay on M6
M5 Pro 64GB (from $1,699)	64GB / 307GB/s	✅ Flies — ~102 tok/s	✅ Unlocked — ~51 tok/s	Q8 quality or max Q4 speed

My take — what I'd actually buy

If you're buying a mini for Qwen3.8-35B-A3B: get the 24GB M6 and run Q4_K_M. It's the value king — fits with KV headroom, fast enough for agentic loops, leaves money for SSD. Pay for 32GB M6 only if you live at 64K+ context. Pay for 64GB M5 Pro only if you've tested Q4 and decided you need Q8's quality bump — Q8 is slower than Q4 on the same chip, so you're trading speed for quality.

Don't buy 16GB for this model. Don't buy M5 Pro for Q4 alone unless you value ~100 tok/s over ~56 tok/s at nearly 2× the price. And don't wait for 35B-A3B to "need less RAM" — the Q4 math is already settled; the weights just need to drop (spotted in ms-swift, not yet on Hugging Face as of Aug 31).

Heads-up — status as of Aug 31, 2026: Qwen3.8-35B-A3B is near-confirmed but not yet officially released (ms-swift table + a Qwen dev "we will be releasing"). All A3B numbers above are from the released Qwen3.6-35B-A3B precursor + the 3.8-27B dense — expect the 3.8-A3B to land within a few % when weights hit Qwen/Qwen3.8-35B-A3B on Hugging Face. I'll update this post when the card goes live.

FAQ — the questions you'll actually search next

Can I run 35B-A3B on a 16GB Mac mini at all?

Technically yes with offload/swap (llama.cpp can spill experts to SSD), but 17.5GB of weights on 16GB unified means you're swapping every token — expect single-digit tok/s and stutter. Not recommended as a daily driver. 24GB is the floor I'd actually use.

Does 27B dense ever beat A3B on a mini?

On raw bench, the dense 27B edges the 3.6-A3B (77.2 vs 73.4 SWE-bench Verified, vendor-reported). In practice on a mini, the A3B's ~9× speed advantage matters more for agentic coding where you run dozens of tool calls — wall-clock wins beat a few bench points. If you do single-shot long reasoning, the dense can justify its slowness.

What about FreeToken / expert offload on a 5090?

For ≤10B-active MoE (A3B's 3B qualifies), a 5090 with FreeToken-style offload can be competitive. But on a Mac mini question, the answer is unified memory vs PCIe: the mini's 24GB unified at 170GB/s has no PCIe penalty, while a 5090 spilling to DDR5 pays it. For this tier, the mini wins on simplicity.

Should I wait for the 35B-A3B weights or buy the mini now?

The mini is the long-lived part — buy on your RAM need (24GB for Q4, 64GB for Q8), not on a weight drop date. The ms-swift commit suggests weeks, not months, but Apple ships Sep 22 — you can order the mini now and pull the model via Unsloth/Ollama the day it lands (Q4 GGUF will be ~18GB download).

Sources: Qwen3.8-27B HF card & README (apache-2.0, 262K ctx), Qwen3.6-35B-A3B release notes (35B/3B MoE), ms-swift supported-models commit (35B-A3B spotted Aug 15), Apple Newsroom Aug 25 2026 (M6 170GB/s, M5 Pro 307GB/s, pricing), plus models-and-hardware-2026.md sizing (Unsloth Q4 ~17–17.5GB, Q8 ~35GB). Benchmark tier: SWE-bench Verified vendor-reported (Qwen 3.6 dense 77.2 vs A3B 73.4). Speed = bandwidth ÷ active-bytes × 0.5, labeled estimate.

Related on this blog: Qwen3.8-27B: 27B Dense, Vision Included, Apache-2.0 · Qwen3.8-Flash-Next & Max coverage

Published Aug 31, 2026. Updated when Qwen/Qwen3.8-35B-A3B weights land. Images are estimated charts (vendor-reported, redrawn). Update promised when Qwen/Qwen3.8-35B-A3B weights land.

Search This Blog

Open Source Factory