What it takes to run the biggest open models
The weights are free to download. The hardware is not. Qwen3.8-2.4T-A95B has 2446B parameters and needs about 2605 GB of memory at Q8_0, the smallest file published so far, at least 36 H100 cards.
Few people run these at a desk, and that isn't the point. It's worth seeing the scale of what “open” means at the top end, and how far it sits from the models on the rest of this section. What your own computer can run →
For scale: the largest Mac on sale, a Mac Studio with M5 Ultra, holds 256GB and lists at $9,499 in that configuration (Apple Store US, checked 22 Oct 2026: $5,499 with 96GB, plus $4,000 for 256GB). That is still 10.2× too small for Qwen3.8-2.4T-A95B.
The models, by size
Two ways to hold them. 9 of these 25 fit a single 256GB Mac Studio, one machine with no server rack. The rest need datacentre GPUs; the last column is how many H100 SXM 80GB cards hold the model.
| Model | Parameters | Memory needed | Fits a 256GB Mac? | H100 cards |
|---|---|---|---|---|
| Qwen3.8-2.4T-A95BAlibaba / Qwen · 95B active | 2446B | 2605 GB | No | 36 |
| DeepSeek-V4-Pro-0813DeepSeek | 1650B | 1786 GB | No | 25 |
| Nex-N2.5-MaxNex-AGI | 1601B | 1658 GB | No | 23 |
| MiMo-V2.6-Pro-RLXiaomi | 1024B | 578 GB | No | 8 |
| Mistral-Large-4-1T-A52BMistral AI · 52B active | 1000B | 1125 GB | No | 16 |
| InklingThinking Machines Lab | 952B | 861 GB | No | 12 |
| Muse-Spark-Open-816B-A42BMeta AI · 42B active | 822B | 1845 GB | No | 26 |
| Hy4-previewTencent | 780B | 471 GB | No | 7 |
| DeepSeek-V4.1-Flash-1017DeepSeek | 763B | 514 GB | No | 8 |
| GLM-5.3Zhipu AI | 753B | 820 GB | No | 12 |
| Atria-Dawn-PreviewShanghai AI Lab | 753B | 1526 GB | No | 21 |
| Atria-DawnShanghai AI Lab | 753B | 1515 GB | No | 21 |
| Step-5-PreviewStepFun | 604B | 732 GB | No | 11 |
| Ling-3.1-flash-1019inclusionAI · 114B active | 572B | 371 GB | No | 6 |
| Ornith-1.5-397BDeepReinforce · 81B active | 403B | 248 GB | Yes | 4 |
| Nex-N2.5-ProNex-AGI · 79B active | 397B | 255 GB | Yes | 4 |
| K2-Horizon-375B-A23BMBZUAI Institute of Foundation Models · 23B active | 379B | 242 GB | Yes | 4 |
| GLM-5.3-FlashZhipu AI | 321B | 385 GB | No | 6 |
| MiMo-V2.6-Flash-RLXiaomi | 311B | 182 GB | Yes | 3 |
| DeepSeek-V4-Flash-Vision-ExpDeepSeek | 305B | 171 GB | Yes | 3 |
| Inkling-SmallThinking Machines Lab | 266B | 166 GB | Yes | 3 |
| Solar-Open2-250BUpstage | 250B | 210 GB | Yes | 3 |
| MiniMax-M3.1-FlashMiniMax · 44B active | 220B | 256 GB | Yes | 4 |
| Qwen3.8-Flash-NextAlibaba / Qwen · 36B active | 180B | 203 GB | Yes | 3 |
| NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16NVIDIA · 12B active | 124B | 282 GB | No | 4 |
Memory is at 4-bit (Q4_K_M) with 4K context; at full precision these models are about four times larger. Models that ship natively quantised are measured from their published weights rather than a conversion. Card counts divide memory by the card's published capacity with a 10% allowance, so they are a floor: real deployments lose more to activations, long-context caches and sharding.
Qwen3.8-2.4T-A95B on different hardware
The same model, and what it would take on each generation of accelerator.
| Accelerator | Memory each | Cards needed |
|---|---|---|
| A100 80GB SXMNVIDIA A100 specification: 80GB HBM2e | 80 GB | 36 |
| H100 SXM 80GBNVIDIA H100 specification: 80GB HBM3 | 80 GB | 36 |
| H200 141GBNVIDIA H200 specification: 141GB HBM3e | 141 GB | 21 |
| B200 180GBNVIDIA DGX B200: 1,440GB across 8 GPUs | 180 GB | 16 |
| Mac Studio with M5 UltraApple sells it with up to 256GB; a 512GB option is announced but not yet on sale | 256 GB | Does not fit |
Card memory is each manufacturer's published capacity per card.
Can't you just quantize it down until it fits?
Not far enough. Quantization has a floor, and these models start so far above it that even the most extreme settings land well beyond any consumer machine. DeepSeek is the example people ask about most, so here are the real published file sizes:
| DeepSeek V3.1, 685B | Download | Fits a 128GB machine? |
|---|---|---|
| Q4_K_M, the normal setting | 405 GB | No |
| Q2_K, already visibly degraded | 246 GB | No |
| UD-IQ1_S, the smallest build in the repository | 192 GB | No |
File sizes from the published GGUF conversions at unsloth/DeepSeek-V3.1-GGUF, checked October 2026. DeepSeek R1's smallest build (unsloth/DeepSeek-R1-GGUF, UD-IQ1_S) is 140 GB, still more than a 128GB machine holds before the context cache is counted.
Squeezing a model to one bit per weight is not free either. By that point it has lost enough precision that it makes mistakes the full version wouldn't, so you are running a noticeably worse model on hardware that is still far beyond a normal computer.
So how is everyone running DeepSeek?
Three ways, and only one of them is a laptop:
- Through an API. Overwhelmingly the most common. The weights being public doesn't mean people host them themselves. It means anyone may, and a handful of providers do.
- On a distilled version. This is the one that causes the confusion. Repos like DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-0528-Qwen3-8B run happily on a laptop and have millions of downloads, but they are Qwen models trained on DeepSeek's output, not DeepSeek's own architecture. When somebody says they're running DeepSeek locally, it is usually one of these.
- On serious hardware. A 256GB Mac Studio or a multi-GPU rig, running a heavily quantized build. Real, but rare.
The same is true of every model on this page. “Open weights” is a licensing fact, not a hardware one.
What about GPT, Claude and Gemini?
We can't tell you. OpenAI, Anthropic and Google don't publish parameter counts or release weights for their frontier models, so any figure for them would be a guess dressed up as a fact. Every number on this page comes from a published file size or an official specification, and we'd rather leave a gap than fill it with invention.
That difference is itself the point of the page. With an open-weight model you can measure exactly how much memory it needs, because the weights are right there. With a closed one you get a price per token and no way to check what sits behind it.