Local models - hardware - verified
Which local LLMs fit your GPU or Mac?
33 GPUs, Macs and AI systems from NVIDIA, AMD and Apple, with memory and bandwidth from each vendor's spec page, and how many of the 63 open-weight models in this directory fit in each one's memory. That runs from 13 models with 8 GB (RTX 5060 Ti 8GB) to 55 with 512 GB (Mac Studio M5 Ultra).
Know your memory but not the model? Browse every local model by memory tier, download size and license.
- Devices
- 33
- Vendors
- 3
- Memory range
- 8-512 GB
- Unified memory
- 10
NVIDIA
16 devicesConsumer GPUs
| Device | Memory | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 | 32 GB | GDDR7 | 1,792 GB/s | 30 / 63 |
| NVIDIA GeForce RTX 5080 | 16 GB | GDDR7 | 960 GB/s | 18 / 63 |
| NVIDIA GeForce RTX 5070 Ti | 16 GB | GDDR7 | 896 GB/s | 18 / 63 |
| NVIDIA GeForce RTX 5070 | 12 GB | GDDR7 | 672 GB/s | 17 / 63 |
| NVIDIA GeForce RTX 5060 Ti 16GB | 16 GB | GDDR7 | 448 GB/s | 18 / 63 |
| NVIDIA GeForce RTX 5060 Ti 8GB | 8 GB | GDDR7 | 448 GB/s | 13 / 63 |
| NVIDIA GeForce RTX 4090 | 24 GB | GDDR6X | 1,008 GB/s | 28 / 63 |
| NVIDIA GeForce RTX 4080 SUPER | 16 GB | GDDR6X | Not stated | 18 / 63 |
| NVIDIA GeForce RTX 4070 Ti SUPER | 16 GB | GDDR6X | Not stated | 18 / 63 |
| NVIDIA GeForce RTX 4060 Ti 16GB | 16 GB | GDDR6 | Not stated | 18 / 63 |
| NVIDIA GeForce RTX 3090 | 24 GB | GDDR6X | Not stated | 28 / 63 |
Desktop AI systems
| Device | Memory (options) | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| NVIDIA DGX Spark | 64 or 128 GB | LPDDR5X unified | 273 GB/s | 42 / 63 |
| NVIDIA DGX Station (GB300) | 252 GB | HBM3E | 7,100 GB/s | 46 / 63 |
Workstation GPUs
| Device | Memory (options) | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | GDDR7 with ECC | 1,792 GB/s | 42 / 63 |
| NVIDIA RTX PRO 5000 Blackwell | 48 or 72 GB | GDDR7 with ECC | 1,344 GB/s | 33 / 63 |
| NVIDIA RTX 6000 Ada Generation | 48 GB | GDDR6 with ECC | 960 GB/s | 32 / 63 |
AMD
9 devicesConsumer GPUs
| Device | Memory | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| AMD Radeon RX 9070 XT | 16 GB | GDDR6 | 640 GB/s | 18 / 63 |
| AMD Radeon RX 9070 | 16 GB | GDDR6 | 640 GB/s | 18 / 63 |
| AMD Radeon RX 7900 XTX | 24 GB | GDDR6 | 960 GB/s | 28 / 63 |
Laptop processors
| Device | Memory | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| AMD Ryzen AI Max+ 395 | 128 GB | LPDDR5x unified (256-bit) | Not stated | 42 / 63 |
Workstation GPUs
| Device | Memory | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| AMD Radeon AI PRO R9700 | 32 GB | GDDR6 | 640 GB/s | 30 / 63 |
| AMD Radeon PRO W7900 | 48 GB | GDDR6 | 864 GB/s | 32 / 63 |
Data center GPUs
| Device | Memory | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| AMD Instinct MI300X | 192 GB | HBM3 | 5,300 GB/s | 43 / 63 |
| AMD Instinct MI325X | 256 GB | HBM3E | 6,000 GB/s | 48 / 63 |
| AMD Instinct MI355X | 288 GB | HBM3E | 8,000 GB/s | 49 / 63 |
Apple
8 devicesMacs
| Device | Memory (options) | Type | Bandwidth | Models that fit |
|---|---|---|---|---|
| MacBook Air (M5, 13-inch and 15-inch) | 16, 24 or 32 GB | Unified memory | 153 GB/s | 30 / 63 |
| MacBook Pro 14-inch (M5) | 16, 24 or 32 GB | Unified memory | 153 GB/s | 30 / 63 |
| MacBook Pro 14-inch and 16-inch (M5 Pro) | 24, 48 or 64 GB | Unified memory | 307 GB/s | 33 / 63 |
| MacBook Pro 14-inch and 16-inch (M5 Max) | 36, 48, 64 or 128 GB | Unified memory | 614 GB/s | 42 / 63 |
| Mac mini (M6) | 16, 24 or 32 GB | Unified memory | 170 GB/s | 30 / 63 |
| Mac mini (M5 Pro) | 24, 48 or 64 GB | Unified memory | 307 GB/s | 33 / 63 |
| Mac Studio (M5 Max) | 36, 48, 64 or 128 GB | Unified memory | 614 GB/s | 42 / 63 |
| Mac Studio (M5 Ultra) | 96, 256 or 512 GB | Unified memory | 1,200 GB/s | 55 / 63 |
Models that fit are counted at each device's largest memory configuration, out of the 63 models in the local models directory. Bandwidth is shown as the vendor states it.
Method
What fits means here
Device specs read on . Model file sizes come from the local models directory, so a device page and the matching memory tier never disagree.
- Specs from the vendor
- Memory, memory type and bandwidth come from NVIDIA, AMD and Apple spec pages, and every device page links the page it was read from. Where a spec page states no bandwidth, the table says Not stated instead of borrowing a figure.
- The 85% rule of thumb
- A model fits when its smallest download the device can run is at most 85% of the device's memory, the same 85% the /local-models/fits pages use. Devices are counted at their largest memory configuration; device pages break it down per configuration.
- MLX counts on Macs only
- MLX builds run on Apple silicon only, so NVIDIA and AMD devices count GGUF, FP8 and official-weight downloads and skip MLX. A Mac counts every format, so its list matches the memory tier page; a GPU page can list a few fewer models and says which ones fit only as MLX.
- Discrete vs unified memory
- On a discrete GPU, fits means the whole download sits in VRAM, with no layers offloaded to system RAM. On unified-memory systems (Macs, DGX Spark, Ryzen AI Max) the operating system and your apps share that memory with the model.
- Nothing about speed
- Fitting in memory says nothing about how fast a model runs. These pages do not measure or estimate speed, and bandwidth is listed as the vendor states it, not converted into a speed figure.
FAQ
How many local models fit on a 24 GB GPU like the RTX 4090?
28 of 63 by this directory's rule of thumb: the smallest download the card can run (GGUF, FP8 or the official weights; MLX builds do not count on a GPU) has to be at most 20.4 GB, 85% of 24 GB. The 24 GB cards in this list are RTX 4090, RTX 3090 and Radeon RX 7900 XTX. The largest by total parameters that fits is Cohere Aya Expanse 32B.
What can an 8 GB or 16 GB graphics card run?
13 models fit in 8 GB and 18 in 16 GB, counting only downloads a graphics card can run. The largest by total parameters are Qwen3.5-9B at 8 GB and gpt-oss-20b at 16 GB. Each device page lists every one with its smallest download and license.
Which device here fits the most models?
The Mac Studio M5 Ultra at up to 512 GB, with 55 of 63. The other 8 have a smallest listed download too big for any single device in this list.
Does unified memory on a Mac count the same as VRAM?
The rule is the same: a model fits when its smallest download is at most 85% of the memory. A Mac can also run MLX builds, which are sometimes the smallest download, so a Mac can list a few more models than a GPU with the same memory. On a Mac, or any unified-memory system, the operating system and your apps share that memory with the model, so the 15% headroom matters more. The 8 Macs here are listed at their largest configuration, and each Mac page shows how the list changes for every configuration Apple sells.
Does fitting in memory mean a model runs fast?
No. These pages only check whether the download fits in memory. They do not measure or estimate speed.
Where do the memory and bandwidth numbers come from?
From NVIDIA, AMD and Apple spec pages, read on October 18, 2026; every device page links its source. The spec pages for 5 devices (RTX 4080 SUPER, RTX 4070 Ti SUPER, RTX 4060 Ti 16GB, RTX 3090 and Ryzen AI Max+ 395) state no bandwidth figure, so those show Not stated rather than a number from elsewhere.
Newsletter
New open weights, every week
One email per week. New model releases, open-weight drops and what they mean for what you build.
Keep exploring
Best local AI models
Our curated picks by hardware tier, with benchmarks and the reasoning.
Model API pricing and specs
First-party prices per 1M tokens when you would rather call an API.
AI benchmark leaderboards
Sourced scores, for when you need to compare quality.
Setup guides
Point your coding tools at the model you picked.