Mac Studio, M3 Ultra, 96GB, refurbished

1TB of storage, 819GB/s of memory bandwidth, and 88GB of unified memory left for a model once macOS has taken its share. It holds 13 of the 22 models we track.

Apple states the original price as $5,299 and the saving as $970 on this build.

Offers

checked 11 minutes ago

Apple Certified Refurbishedcheapest todayin stockchecked 11 minutes ago$4,329
Refurbished by Apple, with Apple's one-year limited warranty and 14-day returns. Apple says refurbished supply is usually very limited and often runs out, so this row can go dead faster than a new one.

No reseller undercuts Apple Certified Refurbished on this Config, so Apple Certified Refurbished direct is the price to beat.

Timing

Collecting price history — 2 of 90 days. No timing verdict yet.

Models it runs — 8K context

Llama-4-Behemoth—
Won't fit— headroom
Kimi-K2-Instruct—
Won't fit— headroom
DeepSeek-V3—
Won't fit— headroom
DeepSeek-R1—
Won't fit— headroom
Qwen3-Coder-480B-A35B-Instruct—
Won't fit— headroom
Llama-4-Maverick-17B-128E—
Won't fit— headroom
GLM-4.6—
Won't fit— headroom
Qwen3-235B-A22B—
Won't fit— headroom
MiniMax-M2—
Won't fit— headroom
Mistral-Large-2411Q4_K_M
Fits10GB headroomest. 7.2 tok/s
gpt-oss-120bQ5_K_M
Tight4GB headroomest. 113 tok/s
Llama-4-Scout-17B-16EQ5_K_M
Fits8GB headroomest. 34.0 tok/s
Llama-3.3-70BQ8_0
Fits10GB headroomest. 7.2 tok/s
Qwen3-32BBF16
Fits21GB headroomest. 8.3 tok/s
Qwen3-30B-A3BBF16
Fits25GB headroomest. 62.0 tok/s
Gemma-3-27BBF16
Fits31GB headroomest. 9.9 tok/s
Devstral-Small-2507BF16
Fits38GB headroomest. 11.1 tok/s
gpt-oss-20bBF16
Fits45GB headroomest. 56.9 tok/s
Phi-4BF16
Fits57GB headroomest. 19.0 tok/s
Qwen3-VL-8B-InstructBF16
Fits70GB headroomest. 33.3 tok/s
Gemma-3n-E4BBF16
Fits73GB headroomest. 66.5 tok/s
Qwen3-4BBF16
Fits78GB headroomest. 66.5 tok/s

Model	Best quant	Fit	Headroom	Est. speed
Llama-4-Behemoth	—	Won't fit	—	—
Kimi-K2-Instruct	—	Won't fit	—	—
DeepSeek-V3	—	Won't fit	—	—
DeepSeek-R1	—	Won't fit	—	—
Qwen3-Coder-480B-A35B-Instruct	—	Won't fit	—	—
Llama-4-Maverick-17B-128E	—	Won't fit	—	—
GLM-4.6	—	Won't fit	—	—
Qwen3-235B-A22B	—	Won't fit	—	—
MiniMax-M2	—	Won't fit	—	—
Mistral-Large-2411	Q4_K_M	Fits	10GB	7.2 tok/s
gpt-oss-120b	Q5_K_M	Tight	4GB	113 tok/s
Llama-4-Scout-17B-16E	Q5_K_M	Fits	8GB	34.0 tok/s
Llama-3.3-70B	Q8_0	Fits	10GB	7.2 tok/s
Qwen3-32B	BF16	Fits	21GB	8.3 tok/s
Qwen3-30B-A3B	BF16	Fits	25GB	62.0 tok/s
Gemma-3-27B	BF16	Fits	31GB	9.9 tok/s
Devstral-Small-2507	BF16	Fits	38GB	11.1 tok/s
gpt-oss-20b	BF16	Fits	45GB	56.9 tok/s
Phi-4	BF16	Fits	57GB	19.0 tok/s
Qwen3-VL-8B-Instruct	BF16	Fits	70GB	33.3 tok/s
Gemma-3n-E4B	BF16	Fits	73GB	66.5 tok/s
Qwen3-4B	BF16	Fits	78GB	66.5 tok/s

Best quant is the highest-quality one that still loads with headroom. Every speed is an estimate; the formula is here.