Mac Studio M5 Ultra 96GB Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the Amazon.jp price reaches your target.
The Mac Studio M5 Ultra 96GB is the entry M5 Ultra model Apple announced in August 2026, with 1.2 TB/s of memory bandwidth, twice that of the M5 Max. It generates about twice as fast on the same model: Gemma 4 31B at 4-bit runs at about 46 tokens/s. Its weak spot is capacity: at 96GB it has less memory than the 128GB M5 Max, so 30B-class models run fast at 8-bit, but the 176B Qwen3.8 Flash Next only fits at 2-bit. Memory can't be added after purchase, so order enough up front.
Amazon.jp
Mac Studio M5 Ultra 96GB price history
Amazon.jp
Some buy links are affiliate links: we may earn a commission if you buy through them, at no extra cost to you.
Full Mac Studio M5 Ultra 96GB specifications
- Memory96 GB · LPDDR5X-9600 · unified memory
- Bandwidth1,229 GB/s
- Chip
- Apple M5 Ultra
- CPU
- 30 / 36-core
- GPU
- 64 / 80-core
- Storage
- 1 TB / 2 TB / 4 TB / 8 TB / 16 TB
- Ecosystem
- Metal
- Power
- 480 W
- Weight
- 3.64 kg
- Model identifier
- Mac17,15
- Part number
- MHL74LL/A
- Type
- Unified memory
- Launch date
- Aug 25, 2026
- MSRP
- $5,499
What LLMs can a single Mac Studio M5 Ultra 96GB run?
The 96GB configuration is best suited to 30B-class models. At 8-bit they fit whole: Gemma 4 31B runs at about 27 tokens/s, and MoE models like Qwen3.6 35B A3B top 120. 4-bit is for speeding up 30B-class models, taking Gemma 4 31B to about 46 tokens/s; no MoE over 100B fits at this tier, as even the smallest, Qwen3.8 Flash Next, needs 111 GB. 2-bit is the only way to run big models on this machine: Qwen3.8 Flash Next at 2-bit is about 79 GB and runs at about 130 tokens/s, while DeepSeek V4 Flash at 2-bit is 96.8 GB and fills nearly the entire machine. macOS lets the GPU use only about three-quarters of memory by default, around 72GB, so both of these 2-bit models need that limit raised by hand first, and don't run other big apps alongside them.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 152 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 70.4 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 82.8 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 161.5 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 68 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 170.2 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | Runs | 130.5 | 256K |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | Runs | 106.1 | 918K |
 Hy3295B (21B active) | 107.2 GB | Too big | — | — |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | Too big | — | — |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 131.6 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 51.2 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 52.1 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 145.5 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 46.3 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 148.3 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | Too big | — | — |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | Too big | — | — |
| Hy3295B (21B active) | 182.2 GB | Too big | — | — |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | Too big | — | — |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 108.7 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 31.2 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 30.8 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 120.4 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 27.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 124.1 | 256K |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | Too big | — | — |
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | Too big | — | — |
| Hy3295B (21B active) | 317.7 GB | Too big | — | — |
| GLM 5.3 Flash320B (18B active) | 341 GB | Too big | — | — |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- You mainly run 30B-class models and want fast generation: Gemma 4 31B at 4-bit runs at about 46 tokens/s, twice the M5 Max
- You want to run your everyday MoE models straight at 8-bit: Qwen3.6 35B A3B at 8-bit runs at about 124 tokens/s, faster than the M5 Max manages at 4-bit
- You often feed it long code or long documents: every GPU core has a neural accelerator, and Apple says prompt processing in LM Studio is up to 4 times as fast as on the M3 Ultra
Not a good fit
- You want MoE models over 100B at 4-bit: Qwen3.8 Flash Next needs 111 GB at 4-bit, which 96GB can't hold, while the 128GB M5 Max can run it
- You want a 300B-class MoE like DeepSeek V4 Flash on one machine: only 2-bit squeezes in, and it fills nearly all the memory; for 4-bit you need an M5 Ultra with 256GB or more
Before you buy
Order enough memory up front
The M5 Ultra starts at 96GB, and the only steps up are 256GB and 512GB, with no 128GB in between. Memory can't be added after purchase, so if you plan to run MoE models over 100B at 4-bit, go straight to 256GB. 512GB is only available with the 36-core CPU, 80-core GPU version, which won't ship until late October.
What the 36-core / 80-core version speeds up
Both chips have 1.2 TB/s of memory bandwidth, so they generate equally fast; the extra 16 GPU cores only speed up prompt processing. Used listings include both, so when inspecting a machine, check the CPU and GPU core counts in System Information and only close the deal once they match what the seller listed.
What it takes to use the neural accelerators
The neural accelerators only kick in with software support: MLX needs macOS 26.2 or later, and official llama.cpp also enables them on macOS 26.2 or later. Users have reported that LM Studio's bundled llama.cpp runtime doesn't enable this path on the M5, making prompt processing for GGUF models 2 to 3 times slower; switching to the MLX engine in LM Studio avoids the problem.
Linking multiple units
The M5 Ultra has 6 Thunderbolt 5 ports (4 on the back, 2 on the front), and starting with macOS 26.2, RDMA over Thunderbolt 5 is supported: link several units into a cluster to pool their memory and run models too big for one. Apple says a four-unit cluster runs inference up to 3 times as fast as a single unit. RDMA is off by default and has to be enabled in Recovery mode.
Clear Activation Lock on a used machine
If Activation Lock is on, an Apple silicon Mac still needs the original owner's Apple Account to activate, even after being erased and reinstalled. Before the deal, have the seller turn off Find My Mac in Find My and sign out of their Apple Account, then power it on in front of you to confirm.
Compared with similar machines
| Model | Amazon.jp median | Memory | Bandwidth | Gemma 4 31B speed t/s | Largest dense model | Power |
|---|---|---|---|---|---|---|
| Mac Studio M5 Ultra 96GBThis page | No price | 96 GB | 1,229 GB/s | 46.3 | 165B | 480 W |
| Mac Studio M5 Max 128GB | No price | 128 GB | 614 GB/s | 23.9 | 222B | 480 W |
| Mac Studio M3 Ultra 512GB | No price | 512 GB | 819 GB/s | 31.5 | 911B | 480 W |
 RTX PRO 6000 | No price | 96 GB | 1,792 GB/s | 65.5 | 165B | 600 W |
vs Mac Studio M5 Max 128GB
Two machines from the same generation, and the trade-off is speed versus capacity. The M5 Ultra has twice the M5 Max's bandwidth, so on Gemma 4 31B at 4-bit it runs at about 46 tokens/s to the M5 Max's 24. But it has only 96GB: Qwen3.8 Flash Next at 4-bit (111 GB) fits on the M5 Max and not on the M5 Ultra, which has to drop to 2-bit. If your everyday models all stay under 70GB and fast generation matters, take the M5 Ultra; for 100B-class MoE models at 4-bit, the 128GB M5 Max is the better fit.
vs Mac Studio M3 Ultra 512GB
The M3 Ultra 512GB wins on capacity: DeepSeek V4 Flash fits whole at 4-bit and runs at about 62 tokens/s, while the 96GB M5 Ultra can only squeeze it in at 2-bit. The M5 Ultra wins on speed: with 50% more bandwidth, it runs Gemma 4 31B at 4-bit at about 46 tokens/s to the M3 Ultra's 32, and its GPU adds neural accelerators, which Apple says make prompt processing up to 4 times as fast. The M3 Ultra 512GB is also no longer sold new, only used or as leftover stock. For 300B-class MoE on one machine, take the M3 Ultra or wait for the 512GB M5 Ultra; if you mainly run 30B-class models, the M5 Ultra is much faster.
vs RTX PRO 6000
Both have 96GB, so they fit almost the same models. The PRO 6000 has more bandwidth, running Gemma 4 31B at 4-bit at about 66 tokens/s to the M5 Ultra's 46, and it has the full CUDA software stack, so serving frameworks like vLLM work out of the box. When a model doesn't fit, it can also put part of it in system RAM and keep running, just much more slowly; the Mac has no such fallback. But the PRO 6000 is only a card, and it still needs a host machine and a high-wattage power supply. For the fastest generation and CUDA, take the PRO 6000; for a complete machine you plug in and run quietly on your desk, take the M5 Ultra.
Other markets
eBay
XianyuAs an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.