Can Qwen3.8-Flash-Next run on Mac Studio M3 Ultra 96GB?

YES — Runs great

IQ1_S · 4K context · formula estimate

What this means

Qwen3.8-Flash-Next fits on the Mac Studio M3 Ultra · 96GB at IQ1_S. We estimate it uses 85.7GB of the conservative 88GB working budget.

Estimated decode speed is 117.7–164.7 tokens/s. A roughly 300-word answer may take around 3 seconds.

View technical details ↓Setup guide →

117.7–164.7tokens/s

estimated · instant

85.7GB

needed at IQ1_S

IQ1_S

quant selected

2.3GB

working headroom

needs 85.7 GBworking budget 88 GB

Apple M3 Ultra96 GB unified memory MetalActive cooling

See how fast it feels

Using the midpoint of our 117.7–164.7 tokens/s estimate for this demo.

Live demo · 141.2 tokens/s

▌

Compare other speeds →

Where the memory goes

Component	Detail	GB
Model weights	IQ1_S GGUF (72.5 GB) + mmap overhead	76.1
KV cache	4K context window	8.8
Runtime	macOS inference app + compute buffers	0.8
Total needed	at IQ1_S, 4K context	85.7
Working budget	96 GB unified memory − conservative macOS reserve	88
Headroom	remaining inside the working budget	2.3

Pick your quant

Quant	Download	Memory	Estimated speed	Verdict
IQ1_S ★	72.5 GB	85.7 GB	~117.7–164.7 tok/s	✓ Runs great
IQ4_XS	93.7 GB	107.9 GB	—	✕ Won't fit

Get it running on this Mac

Choose the beginner-friendly build instead.

This Mac has enough memory, but the tracked IQ1_S ternary file is not supported by the current point-and-click apps. For a first local chat, use Bonsai 27B 1-bit in Locally AI. Do not download PQ2_0; the publisher marks it unsupported.

Only use the publisher tool if you are comfortable with Terminal.

The advanced path is complete below, but it is not required for ordinary local AI.

Advanced publisher setup (Terminal)

Install Apple's command-line tools if macOS asks, then run:

git clone https://github.com/PrismML-Eng/Bonsai-demo.git
cd Bonsai-demo
BONSAI_OPENWEBUI=0 BONSAI_SKIP_MLX=1 BONSAI_FAMILY=undefined BONSAI_MODEL=undefined ./setup.sh
BONSAI_FAMILY=undefined BONSAI_MODEL=undefined BONSAI_CTX=4096 ./scripts/start_llama_server.sh

When the server says it is ready, open http://localhost:8080. Start with 4K context. This route uses the publisher's Prism runtime, not Jan or LM Studio.

Other models on this Mac

Ling 3.0 Flash · Q4_K_M GLM-4.5V · Q4_K_M GLM-4.5-Air · Q4_K_M Llama 3.3 70B · Q4_K_M BigBang V1 36B-A3B · Q4_K_M Qwen 3.5 35B-A3B · Q4_K_M

Qwen3.8-Flash-Next on other Mac Studio M3 Ultra configurations

Check 256GB Check 512GB

FAQ

Can the Mac Studio M3 Ultra · 96GB run Qwen3.8-Flash-Next?

Yes at IQ1_S. We estimate about 85.7GB of working memory and 117.7–164.7 tokens/s at 4K context.

Which Qwen3.8-Flash-Next quant should I use on this Mac?

IQ1_S. It is a 72.5GB download and leaves about 2.3GB inside our conservative working budget.

Which app should I use for Qwen3.8-Flash-Next on this Mac?

Current beginner Mac apps do not load this exact file. Use Bonsai 27B 1-bit in Locally AI instead, or open the advanced publisher instructions on this page.

Are these speeds measured on a Mac Studio M3 Ultra?

No. The range is a formula estimate based on memory bandwidth, model size, active parameters, cooling, and a 4K context. The app, backend, thermals, and prompt can change real performance.