Mini PCs and desktops for local AI
A mini PC can run private AI on your own desk if the AI model fits in the memory the AI can use. This page lists mini PCs and small desktops by memory, as each maker publishes it, with the page we read and the day we read it. We have not tested any of them. Our estimates are labelled as estimates.
How to read this page
- Memory decides what runs. The model file, the document you give it and the rest of the computer all share the memory, so a model larger than the memory the AI can use will run badly or not at all.
- Each table below applies the finder's rules to one machine and one memory option. The results are estimated from the specifications and have not been measured. Speeds are always a range, never one number.
- The speed estimates for Apple's machines use a scale worked out from published results for Apple chips. We do not apply that scale to other makers' machines, so their speeds use the plain rule and are a rougher guess.
- Where a maker does not publish a figure, the table says not known. We do not fill gaps.
The list
| Product | Operating system | Chip | Memory options | Memory type | Largest memory the maker states | Memory can be replaced or added | Memory bandwidth | Sources | Read on |
|---|---|---|---|---|---|---|---|---|---|
| Apple Mac mini (M6) with 16GB 2026 Apple lists 153GB/s with 16GB. Memory soldering not stated. Base name Mac mini (M6) is Apple's (compare page); the memory in the row name is our own label and not Apple's. The support page title is 'M6 or M5 Pro'. |
macOS | Apple M6 | 16 GB | unified memory | not known | not known | 153 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac mini (M6) with 24GB or 32GB 2026 Apple lists 170GB/s with 24GB or 32GB. Memory soldering not stated. Base name Mac mini (M6) is Apple's (compare page); the memory in the row name is our own label and not Apple's. The support page title is 'M6 or M5 Pro'. |
macOS | Apple M6 | 24, 32 GB | unified memory | 32 GB | not known | 170 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac mini (M5 Pro) 2026 307GB/s is listed under the M5 Pro chip on both Apple pages; the configurable 18-core CPU/20-core GPU variant has no separate figure. Memory soldering not stated. Apple lists 48GB or 64GB without saying whether the base chip can take them. |
macOS | Apple M5 Pro | 24, 48, 64 GB | unified memory | 64 GB | not known | 307 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac Studio (M5 Max, 32-core GPU) 2026 Apple lists 460GB/s and 36GB for the M5 Max with the 32-core GPU. Memory soldering not stated. The GPU core count in the row name is our own label and not Apple's; Apple lists it per configuration. |
macOS | Apple M5 Max | 36 GB | unified memory | not known | not known | 460 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac Studio (M5 Max, 40-core GPU) 2026 Apple lists 614GB/s for the M5 Max with the 18-core CPU and 40-core GPU, with 48GB, 64GB or 128GB. Memory soldering not stated. The GPU core count in the row name is our own label and not Apple's; Apple lists it per configuration. |
macOS | Apple M5 Max | 48, 64, 128 GB | unified memory | 128 GB | not known | 614 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac Studio (M5 Ultra) 2026 Apple states 1.2TB/s; entered as 1200 GB/s (TB x 1000). Apple lists 96GB with the 30-core CPU and 64-core GPU configuration; larger memory is on the 80-core GPU row. Memory soldering not stated. |
macOS | Apple M5 Ultra | 96 GB | unified memory | not known | not known | 1200 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| Apple Mac Studio (M5 Ultra, 80-core GPU) 2026 Apple states 1.2TB/s; entered as 1200 GB/s (TB x 1000). Apple lists 256GB and 512GB with the 36-core CPU and 80-core GPU configuration. Memory soldering not stated. The GPU core count in the row name is our own label and not Apple's. |
macOS | Apple M5 Ultra | 256, 512 GB | unified memory | 512 GB | not known | 1200 GB/s | apple.com support.apple.com apple.com |
2026-10-30 |
| NVIDIA DGX Spark Spec table states a 256-bit memory interface and 273 GB/s. Page says the 64GB configuration is available only through OEM partners. Memory soldering not stated. Kind from NVIDIA nav category 'Workstation'. |
DGX OS | NVIDIA GB10 Grace Blackwell Superchip | 64, 128 GB | LPDDR5x, coherent unified system memory | 128 GB | not known | 273 GB/s | nvidia.com | 2026-10-30 |
| Lenovo ThinkStation PGX Lenovo PSREF memory line states 256-bit and 273 GB/s. Page lists a single 128GB configuration. PSREF form factor says Mini-tower; kind from the ThinkStation (workstation) line. Soldering not stated. |
NVIDIA DGX OS | NVIDIA GB10 Grace Blackwell Superchip | 128 GB | LPDDR5x unified system memory | 128 GB | not known | 273 GB/s | psref.lenovo.com | 2026-10-30 |
| Beelink GTR9 Pro AMD Ryzen AI Max+ 395 Beelink option text says memory is soldered; one memory option listed. 256 GB/s: Derived: 256 x 8000 / 8 (AMD page: 256-bit LPDDR5x-8000; Beelink: 8000MT/s); AMD Halo page also states 256GB/s. |
Windows 11 Pro | AMD Ryzen AI Max+ 395 | 128 GB | LPDDR5x-8000 | 128 GB | No | 256 GB/s | bee-link.com amd.com amd.com |
2026-10-30 |
| GMKtec EVO-X2 AMD Ryzen AI Max+ 395 AI Mini PC Page: onboard LPDDR5X (non-upgradeable), options 64GB or 128GB (an LLM table also names 96GB). 256 GB/s: Derived: 256 x 8000 / 8 from AMD page; GMKtec gives speed as 8000MHz. |
Windows 11 Pro; Ubuntu | AMD Ryzen AI Max+ 395 | 64, 128 GB | Onboard LPDDR5X | 128 GB | No | 256 GB/s | gmktec.com amd.com amd.com |
2026-10-30 |
| MINISFORUM MS-S1 MAX 64GB Local AI Pilot Edition Page says the 64GB memory chips are soldered; upgrade to 128GB not possible (shared spec table still says max 128GB). 256 GB/s: Derived: 256 x 8000 / 8 (page: LPDDR5x-256Bit, LPDDR5x-8000MT/s). |
Windows 11 Pro | AMD Ryzen AI Max+ 395 | 64 GB | LPDDR5x-8000MT/s | 64 GB | No | 256 GB/s | store.minisforum.com | 2026-10-30 |
| MINISFORUM MS-S1 MAX 128GB Max AI Compute Edition The maker's product data embedded in this page (not shown in the visible text) says memory is soldered rather than socketed, so capacity is fixed at purchase. 256 GB/s: Derived: 256 x 8000 / 8 (page: LPDDR5x-256Bit, LPDDR5x-8000MT/s). |
Windows 11 Pro | AMD Ryzen AI Max+ 395 | 128 GB | LPDDR5x-8000MT/s | 128 GB | No | 256 GB/s | store.minisforum.com | 2026-10-30 |
| HP Z2 Mini G1a Workstation Desktop PC Singapore regional pages covering one SKU (128 GB, listed as LPDDR5x-8533 MT/s onboard). Other memory options, maximum and soldering are not stated on the pages read. Bandwidth left empty: HP's 8533 differs from AMD's 8000. chip_id empty: PRO 395 is not in chips.csv. |
Windows 11 Pro | AMD Ryzen AI Max+ PRO 395 | 128 GB | LPDDR5x 8533 | not known | not known | not known | hp.com hp.com |
2026-10-30 |
| ASUS NUC 15 Pro Plus Mini PC includes 2x8GB or 2x16GB DDR5-6400 CSO-DIMM (totals computed x2); kit ships without memory; 96GB (2x48GB) is the stated maximum, not a configuration sold. Kit has no OS. One row for three CPUs: chip_id empty. Tool-free RAM upgrade stated. |
Windows 11 Home; Windows 11 Pro | Intel Core Ultra 5 225H; Intel Core Ultra 7 255H; Intel Core Ultra 9 285H | 16, 32 GB | DDR5-5600 SO-DIMM or DDR5-6400 CSO-DIMM | 96 GB | Yes | not known | asus.com asus.com |
2026-10-30 |
| Lenovo ThinkCentre M90q Gen 6 PSREF: up to 64GB (2x 32GB DDR5 SO-DIMM), two SO-DIMM slots; individual sizes and user-replaceability not stated. Page also shows optional discrete Arc A310 graphics, so gpu_type is empty. Form factor Tiny (1L). |
Windows 11 Pro; Windows 11 Home; Windows 11 Home Single Language; Windows 11 IoT Enterprise LTSC 2024 | Intel Core Ultra 5 / 7 / 9 Processor (Series 2) | not known | DDR5-5600 | 64 GB | not known | not known | psref.lenovo.com | 2026-10-30 |
What fits, by memory size
Estimated from specs Worked out from each maker's published memory figures and the size of each model, with the finder's rules. Nothing here was measured on these machines. Speed is the speed of writing an answer, shown as a range. Reading speed is not estimated. How the estimates work · The finder
Apple Mac mini (M6) with 16GB
With 16 GB of memory
About 11.2 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 31 to 57 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 25 to 46 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 19 to 35 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 14 to 26 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 13 to 23 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 12 to 22 | Fast |
| Gemma 3 12B gemma3:12b |
Won't fit | 15.9 | not known | n/a |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b |
Won't fit | 13.7 | not known | n/a |
| Phi-4 14B phi4:14b |
Won't fit | 13.9 | not known | n/a |
| Qwen3 14B qwen3:14b |
Won't fit | 13.4 | not known | n/a |
| gpt-oss 20B gpt-oss:20b |
Won't fit | 16.5 | not known | n/a |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b |
Won't fit | 22.6 | not known | n/a |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac mini (M6) with 24GB or 32GB
With 24 GB of memory
About 16.8 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 34 to 64 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 27 to 51 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 21 to 39 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 16 to 29 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 14 to 26 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 13 to 24 | Fast |
| Gemma 3 12B gemma3:12b |
Tight | 15.9 | 8.4 to 16 | Unclear could be Moderate or Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 7.6 to 14 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 7.5 to 14 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 7.4 to 14 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Tight | 16.5 | 17 to 32 | Fast |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b |
Won't fit | 22.6 | not known | n/a |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 32 GB of memory
About 22.4 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 34 to 64 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 27 to 51 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 21 to 39 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 16 to 29 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 14 to 26 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 13 to 24 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 8.4 to 16 | Unclear could be Moderate or Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 7.6 to 14 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 7.5 to 14 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 7.4 to 14 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 17 to 32 | Fast |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b |
Won't fit | 22.6 | not known | n/a |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac mini (M5 Pro)
With 24 GB of memory
About 16.8 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 54 to 100 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 43 to 80 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 33 to 60 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 24 to 45 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 22 to 41 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 21 to 38 | Fast |
| Gemma 3 12B gemma3:12b |
Tight | 15.9 | 13 to 25 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 12 to 22 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 12 to 22 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 12 to 21 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Tight | 16.5 | 27 to 50 | Fast |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b |
Won't fit | 22.6 | not known | n/a |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 48 GB of memory
About 33.6 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 54 to 100 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 43 to 80 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 33 to 60 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 24 to 45 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 22 to 41 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 21 to 38 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 13 to 25 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 12 to 22 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 12 to 22 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 12 to 21 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 27 to 50 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 6.3 to 12 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 31 to 58 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 5.4 to 10 | Moderate |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 64 GB of memory
About 44.8 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 54 to 100 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 43 to 80 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 33 to 60 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 24 to 45 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 22 to 41 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 21 to 38 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 13 to 25 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 12 to 22 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 12 to 22 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 12 to 21 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 27 to 50 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 6.3 to 12 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 31 to 58 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 5.4 to 10 | Moderate |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac Studio (M5 Max, 32-core GPU)
With 36 GB of memory
About 25.2 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 80 to 150 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 64 to 120 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 49 to 91 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 37 to 68 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 33 to 61 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 31 to 57 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 20 to 37 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 18 to 33 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 18 to 33 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 17 to 32 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 40 to 74 | Fast |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 47 to 87 | Fast |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac Studio (M5 Max, 40-core GPU)
With 48 GB of memory
About 33.6 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 75 to 140 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 60 to 112 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 46 to 85 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 34 to 63 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 31 to 57 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 29 to 54 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 19 to 34 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 17 to 31 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 17 to 31 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 16 to 30 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 37 to 70 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 8.8 to 16 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 44 to 82 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 7.5 to 14 | Unclear could be Moderate or Fast |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 64 GB of memory
About 44.8 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 75 to 140 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 60 to 112 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 46 to 85 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 34 to 63 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 31 to 57 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 29 to 54 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 19 to 34 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 17 to 31 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 17 to 31 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 16 to 30 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 37 to 70 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 8.8 to 16 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 44 to 82 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 7.5 to 14 | Unclear could be Moderate or Fast |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 128 GB of memory
About 89.6 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 75 to 140 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 60 to 112 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 46 to 85 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 34 to 63 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 31 to 57 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 29 to 54 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 19 to 34 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 17 to 31 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 17 to 31 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 16 to 30 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 37 to 70 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 8.8 to 16 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 44 to 82 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 7.5 to 14 | Unclear could be Moderate or Fast |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 3.5 to 6.5 | Unclear could be Slow or Moderate |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 32 to 59 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac Studio (M5 Ultra)
With 96 GB of memory
About 67.2 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 147 to 273 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 118 to 218 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 89 to 165 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 67 to 124 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 60 to 111 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 57 to 105 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 36 to 67 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 33 to 61 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 32 to 60 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 32 to 59 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 73 to 136 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 17 to 32 | Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 86 to 159 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 15 to 27 | Fast |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 6.8 to 13 | Unclear could be Moderate or Fast |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Apple Mac Studio (M5 Ultra, 80-core GPU)
With 256 GB of memory
About 179.2 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 147 to 273 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 118 to 218 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 89 to 165 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 67 to 124 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 60 to 111 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 57 to 105 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 36 to 67 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 33 to 61 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 32 to 60 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 32 to 59 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 73 to 136 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 17 to 32 | Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 86 to 159 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 15 to 27 | Fast |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 6.8 to 13 | Unclear could be Moderate or Fast |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 62 to 116 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 512 GB of memory
About 358.4 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 147 to 273 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 118 to 218 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 89 to 165 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 67 to 124 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 60 to 111 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 57 to 105 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 36 to 67 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 33 to 61 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 32 to 60 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 32 to 59 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 73 to 136 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 17 to 32 | Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 86 to 159 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 15 to 27 | Fast |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 6.8 to 13 | Unclear could be Moderate or Fast |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 62 to 116 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
NVIDIA DGX Spark
With 64 GB of memory
About 44.8 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 48 to 89 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 38 to 71 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 29 to 54 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 22 to 40 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 19 to 36 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 18 to 34 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 12 to 22 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 11 to 20 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 10 to 20 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 10 to 19 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 24 to 44 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.6 to 10 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 28 to 52 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.8 to 8.9 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 128 GB of memory
About 89.6 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 48 to 89 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 38 to 71 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 29 to 54 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 22 to 40 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 19 to 36 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 18 to 34 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 12 to 22 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 11 to 20 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 10 to 20 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 10 to 19 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 24 to 44 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.6 to 10 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 28 to 52 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.8 to 8.9 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.2 to 4.1 | Slow |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 20 to 38 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Lenovo ThinkStation PGX
With 128 GB of memory
About 89.6 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 48 to 89 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 38 to 71 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 29 to 54 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 22 to 40 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 19 to 36 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 18 to 34 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 12 to 22 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 11 to 20 | Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 10 to 20 | Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 10 to 19 | Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 24 to 44 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.6 to 10 | Unclear could be Moderate or Fast |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 28 to 52 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.8 to 8.9 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.2 to 4.1 | Slow |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 20 to 38 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Beelink GTR9 Pro AMD Ryzen AI Max+ 395
With 128 GB of memory
About 124 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 45 to 83 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 36 to 67 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 27 to 50 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 20 to 38 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 18 to 34 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 17 to 32 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 11 to 21 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 10 to 18 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 9.8 to 18 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 9.6 to 18 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 22 to 41 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.3 to 9.8 | Moderate |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 26 to 49 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.5 to 8.3 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.1 to 3.9 | Slow |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 19 to 35 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
GMKtec EVO-X2 AMD Ryzen AI Max+ 395 AI Mini PC
With 64 GB of memory
About 60 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 45 to 83 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 36 to 67 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 27 to 50 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 20 to 38 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 18 to 34 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 17 to 32 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 11 to 21 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 10 to 18 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 9.8 to 18 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 9.6 to 18 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 22 to 41 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.3 to 9.8 | Moderate |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 26 to 49 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.5 to 8.3 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.1 to 3.9 | Slow |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 128 GB of memory
About 124 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 45 to 83 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 36 to 67 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 27 to 50 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 20 to 38 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 18 to 34 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 17 to 32 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 11 to 21 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 10 to 18 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 9.8 to 18 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 9.6 to 18 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 22 to 41 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.3 to 9.8 | Moderate |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 26 to 49 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.5 to 8.3 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.1 to 3.9 | Slow |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 19 to 35 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
MINISFORUM MS-S1 MAX 64GB Local AI Pilot Edition
With 64 GB of memory
About 60 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 45 to 83 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 36 to 67 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 27 to 50 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 20 to 38 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 18 to 34 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 17 to 32 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 11 to 21 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 10 to 18 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 9.8 to 18 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 9.6 to 18 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 22 to 41 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.3 to 9.8 | Moderate |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 26 to 49 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.5 to 8.3 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.1 to 3.9 | Slow |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
MINISFORUM MS-S1 MAX 128GB Max AI Compute Edition
With 128 GB of memory
About 124 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 45 to 83 | Fast |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 36 to 67 | Fast |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 27 to 50 | Fast |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 20 to 38 | Fast |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 18 to 34 | Fast |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 17 to 32 | Fast |
| Gemma 3 12B gemma3:12b |
Fits | 15.9 | 11 to 21 | Fast |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first |
Fits | 13.7 | 10 to 18 | Unclear could be Moderate or Fast |
| Phi-4 14B phi4:14b |
Fits | 13.9 | 9.8 to 18 | Unclear could be Moderate or Fast |
| Qwen3 14B qwen3:14b · thinks first |
Fits | 13.4 | 9.6 to 18 | Unclear could be Moderate or Fast |
| gpt-oss 20B gpt-oss:20b · thinks first |
Fits | 16.5 | 22 to 41 | Fast |
| Gemma 3 27B gemma3:27b |
Fits | 27.2 | 5.3 to 9.8 | Moderate |
| Qwen3 30B (A3B) qwen3:30b · thinks first |
Fits | 22.6 | 26 to 49 | Fast |
| Qwen3 32B qwen3:32b · thinks first |
Fits | 26.3 | 4.5 to 8.3 | Unclear could be Slow or Moderate |
| Llama 3.1 70B llama3.1:70b |
Fits | 51.5 | 2.1 to 3.9 | Slow |
| gpt-oss 120B gpt-oss:120b · thinks first |
Fits | 70.5 | 19 to 35 | Fast |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
HP Z2 Mini G1a Workstation Desktop PC
Speed is not known for this machine. The maker does not publish a memory bandwidth figure we could source, so the tables show fit only.
With 128 GB of memory
About 124 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 5 | not known | Not known |
| Qwen3 4B qwen3:4b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6 | not known | Not known |
| Gemma 3 4B gemma3:4b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6.7 | not known | Not known |
| Mistral 7B v0.3 mistral:7b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 7.8 | not known | Not known |
| Llama 3.1 8B llama3.1:8b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.3 | not known | Not known |
| Qwen3 8B qwen3:8b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.9 | not known | Not known |
| Gemma 3 12B gemma3:12b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 15.9 | not known | Not known |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.7 | not known | Not known |
| Phi-4 14B phi4:14b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.9 | not known | Not known |
| Qwen3 14B qwen3:14b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.4 | not known | Not known |
| gpt-oss 20B gpt-oss:20b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 16.5 | not known | Not known |
| Gemma 3 27B gemma3:27b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 27.2 | not known | Not known |
| Qwen3 30B (A3B) qwen3:30b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 22.6 | not known | Not known |
| Qwen3 32B qwen3:32b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 26.3 | not known | Not known |
| Llama 3.1 70B llama3.1:70b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 51.5 | not known | Not known |
| gpt-oss 120B gpt-oss:120b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 70.5 | not known | Not known |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
ASUS NUC 15 Pro Plus
Speed is not known for this machine. The maker does not publish a memory bandwidth figure we could source, so the tables show fit only.
With 16 GB of memory
About 12 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 5 | not known | Not known |
| Qwen3 4B qwen3:4b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6 | not known | Not known |
| Gemma 3 4B gemma3:4b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6.7 | not known | Not known |
| Mistral 7B v0.3 mistral:7b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 7.8 | not known | Not known |
| Llama 3.1 8B llama3.1:8b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.3 | not known | Not known |
| Qwen3 8B qwen3:8b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.9 | not known | Not known |
| Gemma 3 12B gemma3:12b |
Won't fit | 15.9 | not known | n/a |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b |
Won't fit | 13.7 | not known | n/a |
| Phi-4 14B phi4:14b |
Won't fit | 13.9 | not known | n/a |
| Qwen3 14B qwen3:14b |
Won't fit | 13.4 | not known | n/a |
| gpt-oss 20B gpt-oss:20b |
Won't fit | 16.5 | not known | n/a |
| Gemma 3 27B gemma3:27b |
Won't fit | 27.2 | not known | n/a |
| Qwen3 30B (A3B) qwen3:30b |
Won't fit | 22.6 | not known | n/a |
| Qwen3 32B qwen3:32b |
Won't fit | 26.3 | not known | n/a |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
With 32 GB of memory
About 28 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed |
|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 5 | not known | Not known |
| Qwen3 4B qwen3:4b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6 | not known | Not known |
| Gemma 3 4B gemma3:4b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 6.7 | not known | Not known |
| Mistral 7B v0.3 mistral:7b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 7.8 | not known | Not known |
| Llama 3.1 8B llama3.1:8b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.3 | not known | Not known |
| Qwen3 8B qwen3:8b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 8.9 | not known | Not known |
| Gemma 3 12B gemma3:12b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 15.9 | not known | Not known |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.7 | not known | Not known |
| Phi-4 14B phi4:14b Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.9 | not known | Not known |
| Qwen3 14B qwen3:14b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 13.4 | not known | Not known |
| gpt-oss 20B gpt-oss:20b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 16.5 | not known | Not known |
| Gemma 3 27B gemma3:27b Speed is not known: no sourced memory bandwidth for this machine. |
Tight | 27.2 | not known | Not known |
| Qwen3 30B (A3B) qwen3:30b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Fits | 22.6 | not known | Not known |
| Qwen3 32B qwen3:32b · thinks first Speed is not known: no sourced memory bandwidth for this machine. |
Tight | 26.3 | not known | Not known |
| Llama 3.1 70B llama3.1:70b |
Won't fit | 51.5 | not known | n/a |
| gpt-oss 120B gpt-oss:120b |
Won't fit | 70.5 | not known | n/a |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Lenovo ThinkCentre M90q Gen 6
The memory type or the memory options are not published, so the finder cannot estimate this one.
Questions to ask before you buy
Ask for the answers in writing, from the maker's own page where you can.
- Exactly which version is this, and does its memory differ between versions?
- How much memory does it have? Can you add or replace it, and where does the maker say so?
- Is the memory shared with the graphics, or does the graphics chip have its own? How much of it can the AI use?
- Does your AI software work with this machine's graphics?
- What are the warranty and return terms, in writing?
What this page leaves out
- Nothing on this page was tested by us. The numbers are the makers' published specifications and the speeds are estimates made from them.
- No prices are shown. Prices change often and differ by country, so check the seller.
- The list holds only the machines we have read so far. A machine that is missing is not a bad machine.
- A maker can change a product after we read its page. Each row shows the day we read it.