Requirements · Ollama and LM Studio
Ollama and LM Studio hardware requirements
Ollama and LM Studio need about 7 GB of VRAM or unified memory for a 9B model at 4-bit, ~13 GB for a 20B model and ~46 GB for a 70B model. A GPU isn’t required, but memory bandwidth sets the speed. Find each model’s minimum below.
Minimum memory for each model in Ollama
| Model | 4-bit | 8-bit | Smallest GPU | Smallest unified | Cheapest tracked rig |
|---|---|---|---|---|---|
| Qwen3.5 9Bchat, small agents | ~7 GB | ~12 GB | 8 GB | 16 GB | Mac mini M5 Pro$1,699 |
| Gemma 4 12Bchat, vision | ~9 GB | ~15 GB | 12 GB | 16 GB | Mac mini M5 Pro$1,699 |
| GPT-OSS 20Breasoning, tools | ~13 GB | ~25 GB | 16 GB | 24 GB | Mac mini M5 Pro$1,699 |
| Gemma 4 26B-A4Bfast general chat | ~18 GB | ~30 GB | 24 GB | 24 GB | GMKtec EVO-X2 64GB$2,200 |
| Qwen3.8 27Bcoding, vision, general | ~19 GB | ~32 GB | 24 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Gemma 3 27Bgeneral, vision | ~19 GB | ~32 GB | 24 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| GLM-4.7-Flash 30B-A3Bcoding agents | ~21 GB | ~35 GB | 24 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Qwen3 30B-A3Bfast chat & coding | ~21 GB | ~36 GB | 24 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Gemma 4 31Bgeneral, vision | ~21 GB | ~36 GB | 24 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Qwen3 32Bgeneral, coding, reasoning | ~22 GB | ~38 GB | 32 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| DeepSeek R1 Distill Qwen 32Breasoning | ~22 GB | ~38 GB | 32 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Qwen3.6 35B-A3Bfast coding & chat | ~24 GB | ~41 GB | 32 GB | 32 GB | GMKtec EVO-X2 64GB$2,200 |
| Llama 3.3 70Bgeneral | ~46 GB | ~80 GB | 96 GB | 64 GB | GMKtec EVO-X2 64GB$2,200 |
| DeepSeek R1 Distill Llama 70Breasoning | ~46 GB | ~80 GB | 96 GB | 64 GB | GMKtec EVO-X2 64GB$2,200 |
| Qwen3-Next 80B-A3Blong context, fast | ~52 GB | ~91 GB | 96 GB | 96 GB | GMKtec EVO-X2 128GB$3,500 |
| Llama 4 Scout 109B-A17Bmultimodal, long context | ~71 GB | ~123 GB | 96 GB | 96 GB | GMKtec EVO-X2 128GB$3,500 |
| GPT-OSS 120Breasoning, tools | ~67 GB | ~132 GB | 96 GB | 96 GB | GMKtec EVO-X2 128GB$3,500 |
| Mistral Small 4 119B-A6Bchat, reasoning, coding, vision | ~78 GB | ~135 GB | 96 GB | 128 GB | GMKtec EVO-X2 128GB$3,500 |
| Nemotron 3 Super 120B-A12Bagents | ~78 GB | ~135 GB | 96 GB | 128 GB | GMKtec EVO-X2 128GB$3,500 |
| Qwen3.5 122B-A10Bstrong general / coding | ~79 GB | ~138 GB | 96 GB | 128 GB | GMKtec EVO-X2 128GB$3,500 |
| Qwen3 235B-A22Bfrontier-class open model | ~151 GB | ~264 GB | multi-GPU | 256 GB | — |
| Muse Spark Open 816B-A42Bcoding, long-horizon agents | ~521 GB | ~912 GB | multi-GPU | — | — |
Smallest standard VRAM size (8, 12, 16, 24, 32, 48, 96 GB) with ~93% usable, and smallest unified-memory size with ~75% usable, that hold the 4-bit model at ~8K context.
How fast will Ollama run on each kind of hardware?
One representative rig per family (the cheapest priced, or the Ultra for Mac Studio).
| Model | Mac mini M5 Pro | Mac Studio M5 Ultra 96GB | GMKtec EVO-X2 64GB | Lenovo ThinkStation PGX | Gigabyte GeForce RTX 4090 Windforce V2 | NVIDIA GeForce RTX 5090 Founders Edition | PNY NVIDIA RTX PRO 6000 Blackwell |
|---|---|---|---|---|---|---|---|
| Qwen3.5 9B~7 GB at Q4_K_M | Fits~48 tok/s | Fits~139 tok/s | Fits~37 tok/s | Fits~39 tok/s | Fits~133 tok/s | Fits~214 tok/s | Fits~214 tok/s |
| GPT-OSS 20B~13 GB at MXFP4 (native) | Fits~82 tok/s | Fits~119 tok/s | Fits~69 tok/s | Fits~80 tok/s | Fits~209 tok/s | Fits~264 tok/s | Fits~264 tok/s |
| Qwen3.6 35B-A3B~24 GB at Q4_K_M | Won't fit | Fits~120 tok/s | Fits~72 tok/s | Fits~83 tok/s | Won't fit | Fits~269 tok/s | Fits~269 tok/s |
| Qwen3.8 27B~19 GB at Q4_K_M | Won't fit | Fits~55 tok/s | Fits~12 tok/s | Fits~13 tok/s | Fits~49 tok/s | Fits~83 tok/s | Fits~83 tok/s |
| Llama 3.3 70B~46 GB at Q4_K_M | Won't fit | Fits~23 tok/s | Tight fit~5 tok/s | Fits~5 tok/s | Won't fit | Won't fit | Fits~34 tok/s |
| GPT-OSS 120B~67 GB at MXFP4 (native) | Won't fit | Tight fit~109 tok/s | Won't fit | Fits~61 tok/s | Won't fit | Won't fit | Fits~231 tok/s |
Estimated single-user decode at ~8K context; GPU-card figures assume the model is fully in VRAM. How we estimate
Does Ollama need a GPU?
No, Ollama and LM Studio both run on CPU only, but generation is limited by memory bandwidth, and ordinary dual-channel desktop RAM is several times slower than a GPU or Apple/Strix Halo unified memory. A GPU or a unified-memory machine is what makes 20B+ models usable.
Ollama’s default downloads are 4-bit (Q4_K_M) for most models, which is what our 4-bit column assumes. Qwen3.5 9B needs ~7 GB, GPT-OSS 20B ~13 GB, Llama 3.3 70B ~46 GB and GPT-OSS 120B ~67 GB.
LM Studio
LM Studio uses the same llama.cpp engine (plus MLX on Macs), so memory needs are the same. On Apple Silicon, it can also run MLX builds instead of GGUF.
If a model doesn’t fit in VRAM, Ollama splits layers between GPU and system RAM; it still runs, but speed drops toward CPU speed. Our tables assume everything is in VRAM or unified memory.
Measured Ollama runs
Public single-user decode results we calibrate against. Runs differ in build and settings, so compare like with like.
| Hardware | Model / run | Measured tok/s | Our formula | Source |
|---|---|---|---|---|
| Strix Halo256 GB/s | Qwen3.6 35B-A3B Q4_K_MOllama/llama.cpp | 63.0 | 71.5 | source |
| Framework Desktop (Strix Halo)256 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 5.0 | 4.8 | source |
| Dell Pro Max GB10273 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 4.7 | 4.9 | source |
| M3 Ultra 512GB819 GB/s | Llama 3.1 70B Q4_K_MOllama (Geerling) | 14.1 | 15.7 | source |
Frequently asked questions
What are the minimum requirements for Ollama?
Ollama itself is light; the model sets the requirement. A 4-bit Qwen3.5 9B needs ~7 GB of VRAM or unified memory, so an 8–12 GB GPU or a 16 GB Mac is the practical minimum for 7B–9B models.
Can Ollama run without a GPU?
Yes, on CPU only, but slowly: typical dual-channel desktop RAM has a fraction of a GPU's bandwidth, so generation is much slower than on a GPU or a unified-memory machine.
How much VRAM does Ollama need for Llama 70B?
About 46 GB at the default 4-bit quantization: a 96 GB card, two 32 GB GPUs, or a 64 GB+ unified-memory machine.
Is LM Studio faster than Ollama?
They share the llama.cpp engine, so on the same GGUF file speeds are close. On Macs, LM Studio can also run MLX models; see the measured MLX and llama.cpp rows in our benchmarks.
Does Ollama use unified memory on a Mac or Strix Halo?
Yes. On Macs it uses Metal and the GPU can address roughly 75% of unified memory by default; on Strix Halo the usable share depends on the BIOS/driver allocation.