Apple · M3 Ultra · Desktop
Best AI Models for Mac Studio (M3 Ultra, 96GB)
Runs 27–31B and 120B MoE models well; dense 70B fits but is slow at 12 tok/s.
| # | Model | Speed | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 28.1 t/s | 28.1 tok/s | Q4_K_M · 21.7 GB | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 92.7 t/s | 92.7 tok/s | Q4_K_M · 16.5 GB | |
| 5 | Q4_K_M · 19.3 GB · 77.4 t/s | 77.4 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Coding Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 28.1 t/s | 28.1 tok/s | Q4_K_M · 21.7 GB | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 76.6 t/s | 76.6 tok/s | Q4_K_M · 22.0 GB | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 92.7 t/s | 92.7 tok/s | Q4_K_M · 16.5 GB | |
| 5 | Q4_K_M · 19.3 GB · 77.4 t/s | 77.4 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Reasoning Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 28.1 t/s | 28.1 tok/s | Q4_K_M · 21.7 GB | |
| 3 | Qwen3.5 27BAlibaba Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 4 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 92.7 t/s | 92.7 tok/s | Q4_K_M · 16.5 GB | |
| 5 | Q4_K_M · 19.3 GB · 77.4 t/s | 77.4 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Math Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 2 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 32.9 t/s | 32.9 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 76.6 t/s | 76.6 tok/s | Q4_K_M · 22.0 GB | |
| 4 | Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 28.1 t/s | 28.1 tok/s | overall | Q4_K_M · 21.7 GB |
| 5 | Q4_K_M · 19.3 GB · 77.4 t/s | 77.4 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 176 t/s | 176 tok/s | Q4_K_M · 3.4 GB | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 107 t/s | 107 tok/s | Q4_K_M · 13.1 GB | |
| 3 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 92.7 t/s | 92.7 tok/s | Q4_K_M · 16.5 GB | |
| 4 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 90.1 t/s | 90.1 tok/s | Q4_K_M · 6.6 GB | |
| 5 | Q4_K_M · 19.3 GB · 77.4 t/s | 77.4 tok/s | Q4_K_M · 19.3 GB |
The Mac Studio (M3 Ultra, 96GB) is the fastest 96 GB machine: it runs 27–31B and 120B MoE models well. macOS keeps about 3.5 GB, leaving 92.5 GB for models. With 43 TFLOPS it reads long prompts faster than most 96 GB machines: a 4K-token prompt takes 26 s. At its $3,999 MSRP it gives 19.2 tok/s per $1,000, 3rd of the 7 machines compared here. M3 Ultra generation, released 2025.
Mac Studio (M3 Ultra, 96GB) strengths and weaknesses
Strengths
- Runs 120B MoE models: GPT OSS 120B at 75.5 tok/s in 70.8 GB
- Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 74.6 GB free
Weaknesses
- Slow prompt reading: A 32K-token prompt on Qwen3.8 27B waits 3.5 min for the first token
- No CUDA: vLLM, AWQ, GPTQ and FP8 builds don't run; use GGUF or MLX
Memory for models
92.5 GB of 96 GBEverything on the size ladder fits. Two 27B-class models together use 35.7 GB and leave 56.8 GB free.
Speed per dollar and per watt
On Qwen3.6 35B A3B (35B params, 3B active) at Q4_K_M. New means a shop price, either a street quote or the maker's own price; used is compared with used. The last two columns are the speed you get per $1,000 of price and per 100 W of power; higher is better. The Mac Studio (M3 Ultra, 96GB) has no current new price, so it is shown at its MSRP.
| Machine | New price | tok/s | tok/s per $1,000 | tok/s per 100 W |
|---|---|---|---|---|
RTX 509032 GB | about $5,376 | 213 | 39.7 | 37.1 |
| about $1,999 | 46.7 | 23.4 | 33.4 | |
Mac Studio M3 Ultra96 GB | about $3,999 | 76.6 | 19.2 | 28.4 |
Mac Studio M2 Ultra192 GB | about $5,599 | 75.9 | 13.6 | 25.7 |
DGX Spark128 GB | about $3,999 | 54.0 | 13.5 | 22.5 |
Mac Studio M3 Ultra256 GB | about $7,099 | 76.6 | 10.8 | 28.4 |
Mac Studio M3 Ultra512 GB | about $9,499 | 76.6 | 8.1 | 28.4 |
| Machine | Used price | tok/s | tok/s per $1,000 | tok/s per 100 W |
|---|---|---|---|---|
RTX 409024 GBBest used value | about $3,000 | 151 | 50.3 | 33.6 |
RTX 509032 GB | about $6,400 | 213 | 33.3 | 37.1 |
Mac Studio M3 Ultra96 GB | about $4,525 | 76.6 | 16.9 | 28.4 |
Cost per token: the Mac Studio (M3 Ultra, 96GB) against 24–512 GB machines and GPUs
On Qwen3.6 35B A3B (35B params, 3B active, Q4_K_M, 4K context). Speed, cost per token and power per token depend on the model.
- Machine, MSRP
- $3.31
- Power
- $0.15
Renting costs as much as buying after ≈2,700 hours (≈1.9 years at 4 h a day).
Each dot is a machine or GPU; hover or tap it for its numbers, click to open it. Machines without a price for the chosen basis are left out. How we calculate cost per token
Other memory sizes of the same machine first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.6 35B A3B.
Output speed depends on memory bandwidth (819 GB/s). Grey rows need more than the 92.5 GB available.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB90.1tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB60.3tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB32.9tok/s
- 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB76.6tok/s
- 70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB12.3tok/s
- 120B MoEGPT OSS 120B · Q4_K_M · 70.8 GB75.5tok/s
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
Ultra generations: writing vs compute
M1 Ultra → M3 UltraWriting speed follows memory bandwidth and barely moved: 75.9 to 76.6 tok/s since the M1 Ultra. Compute, which sets how fast prompts are read, grew 1.4×. Indexed to the M1 Ultra; speeds on Qwen3.6 35B A3B, same 128 GB.
All 1267 models on the Mac Studio (M3 Ultra, 96GB)
Featured onlyBest quant that fits in 92.5 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
Xglm 564MChat | Q4_K_M | 0.4 GB | 1549.5 t/s | 2K | S100, Runs great | |
| Q4_K_M | 5.7 GB | 116.8 t/s | 8K | S94, Runs great | ||
Llama 160MChat | Q4_K_M | 0.5 GB | 1219.8 t/s | 2K | S100, Runs great | |
NuExtractChat | Q4_K_M | 4.2 GB | 168.6 t/s | 4K | S98, Runs great | |
Granite 3.1 3B A800m InstructChatFunctions | Q4_K_M | 3.4 GB | 132.9 t/s | 131K | S95, Runs great | |
| Q4_K_M | 74.7 GB | 31.9 t/s | 1049K | A70, Runs well | ||
Codegen 350M MonoChatCode | FP16 | 0.8 GB | 744.5 t/s | 2K | S100, Runs great | |
Llama 7BChat | Q4_K_M | 4.5 GB | 128.8 t/s | 2K | S95, Runs great | |
Qwen2 VL 2B InstructVision | Q4_K_M | 2.1 GB | 341.3 t/s | 33K | S100, Runs great | |
| Q4_K_M | 0.3 GB | 1737.3 t/s | 2K | S100, Runs great | ||
Tiny Qwen3ForCausalLMChatFunctions | Q4_K_M | 0.6 GB | 1911.0 t/s | 41K | S100, Runs great | |
Qwen2.5 14B InstructChatFunctions | Q4_K_M | 12.4 GB | 60.0 t/s | 33K | A84, Runs well | |
| Q4_K_M | 6.2 GB | 106.4 t/s | 8K | S92, Runs great | ||
Mixtral 8x22B Instruct v0.1ChatFunctions | Q4_K_M | 88.4 GB | 19.0 t/s | 66K | C35, Usable | |
| Q4_K_M | 6.2 GB | 106.4 t/s | 8K | S92, Runs great |
Specs
- Chip
- M3 Ultra
- CPU cores
- 28
- GPU cores
- 60
- Unified memory
- 96 GB LPDDR5
- For models
- 92.5 GB (macOS keeps 3.5 GB)
- Memory bandwidth
- 819 GB/s
- FP16 (dense)
- 43 TFLOPS
- Neural Engine
- 36 TOPS
- NPU
- 32-core Neural Engine
- Power
- 270 W max
- Size
- Mac Studio
- Released
- 2025-03-12
- MSRP
- $3,999
Mac Studio (M3 Ultra, 96GB) FAQ
Is the Mac Studio (M3 Ultra, 96GB) worth it for local LLMs in 2026?
At its $3,999 MSRP it gives 19.2 tok/s per $1,000 on Qwen3.6 35B A3B, 3rd of the 7 machines compared here. It runs 27–31B and 120B MoE models well; dense 70B fits but is slow at 12 tok/s.
How much memory is available for AI on the Mac Studio (M3 Ultra, 96GB)?
About 92.5 GB of the 96 GB. macOS keeps roughly 3.5 GB for itself and other apps.
What is the largest model the Mac Studio (M3 Ultra, 96GB) can run?
Mistral Medium 3.5 128B (127.7B) is the largest at 4-bit or better: at Q4_K_M it uses 82.8 GB, 90% of the available memory, and runs at about 7.4 tok/s.
Can the Mac Studio (M3 Ultra, 96GB) run Qwen3.8 27B?
Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of the 92.5 GB available, and it is the best model that runs well on this machine.
How long does the Mac Studio (M3 Ultra, 96GB) take to read a long prompt?
On Qwen3.8 27B, about 26 s for a 4K-token prompt and 3.5 min for a 32K-token document before the first word appears.
Is renting a GPU cheaper than buying the Mac Studio (M3 Ultra, 96GB)?
An RTX PRO 6000 Blackwell Workstation Edition (96 GB) rents from $1.47/hr. Renting costs as much as buying after ≈2,700 hours (≈1.9 years at 4 h a day).