- Services
Every open model in our directory, ranked by how well it runs on Google Cloud TPU v5p (95GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
Google Cloud TPU v5p has 95 GB of VRAM, enough to run 264 of the 298 open models we track at 4-bit. The largest that still fits well is Kimi K2 Instruct (1000B), and the top-rated is Hy3 at about 40 tok/s.
264 of 298 models run comfortably on Google Cloud TPU v5p.
See full Google Cloud TPU v5p specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| Hy3Tencent | SS | 40.0 tok/s | 55.7 GB |
| GLM-4.5Z.ai | SS | 42.9 tok/s | 51.8 GB |
| GLM-4.7Z.ai | SS | 42.3 tok/s | 52.6 GB |
| Kimi K2 InstructMoonshot AI | SS | 42.9 tok/s | 51.8 GB |
| Qwen3.5-397B-A17BAlibaba | SS | 48.4 tok/s | 46.0 GB |
| Llama 3 70B InstructMeta | SS | 48.7 tok/s | 45.7 GB |
| Qwen 3.5 OmniAlibaba | SS | 49.3 tok/s | 45.2 GB |
| Apertus 70BEPFL, ETH Zurich, CSCS | SS | 51.3 tok/s | 43.4 GB |
| Llama 2 70B ChatMeta | SS | 51.3 tok/s | 43.4 GB |
| Mixtral 8x22B InstructMistral AI | SS | 51.1 tok/s | 43.6 GB |
| Muse Glimmer-30BMeta | SS | 45.1 tok/s | 49.4 GB |
| Qwen3-32BAlibaba | SS | 41.3 tok/s | 53.9 GB |
Hy3Tencent | 295B(21B active) | SS | 40.0 tok/s | 55.7 GB | |
GLM-4.5Z.ai | 355B(32B active) | SS | 42.9 tok/s | 51.8 GB | |
GLM-4.7Z.ai | 358B(32B active) | SS | 42.3 tok/s | 52.6 GB | |
Kimi K2 InstructMoonshot AI | 1000B(32B active) | SS | 42.9 tok/s | 51.8 GB | |
Qwen3.5-397B-A17BAlibaba | 397B(17B active) | SS | 48.4 tok/s | 46.0 GB | |
| 70B | SS | 48.7 tok/s | 45.7 GB | ||
Qwen 3.5 OmniAlibaba | 397B(17B active) | SS | 49.3 tok/s | 45.2 GB | |
Apertus 70BEPFL, ETH Zurich, CSCS | 70B | SS | 51.3 tok/s | 43.4 GB | |
Ad Tools for AI BuildersAll the hardware and tools you need to run private AI at your office or at home.Shop Now | |||||
Llama 2 70B ChatMeta | 70B | SS | 51.3 tok/s | 43.4 GB | |
Mixtral 8x22B InstructMistral AI | 141B(39B active) | SS | 51.1 tok/s | 43.6 GB | |
Muse Glimmer-30BMeta | 30B | SS | 45.1 tok/s | 49.4 GB | |
Qwen3-32BAlibaba | 32.8B | SS | 41.3 tok/s | 53.9 GB | |
Gemma 3 27B ITGoogle | 27B | SS | 50.8 tok/s | 43.8 GB | |
DeepSeek-V3DeepSeek | 671B(37B active) | SS | 37.2 tok/s | 59.8 GB | |
DeepSeek-R1DeepSeek | 671B(37B active) | SS | 37.2 tok/s | 59.8 GB | |
DeepSeek-V3.1DeepSeek | 671B(37B active) | SS | 37.2 tok/s | 59.8 GB | |
Ad Vast.aiAffordable on-demand GPU rentals for training and inference. Pick from thousands of hosts.Rent a GPU | |||||
DeepSeek-V3.2DeepSeek | 685B(37B active) | SS | 37.2 tok/s | 59.8 GB | |
Qwen3-235B-A22BAlibaba | 235B(22B active) | SS | 61.3 tok/s | 36.3 GB | |
Mistral Small 3 24BMistral AI | 24B | SS | 57.1 tok/s | 39.0 GB | |
LLaMA 65BMeta | 65B | SS | 56.7 tok/s | 39.3 GB | |
Qwen3.5-122B-A10BAlibaba | 122B(10B active) | SS | 81.6 tok/s | 27.3 GB | |
KolibriAleph Alpha | 78B(3B active) | SS | 84.8 tok/s | 26.2 GB | |
Mistral Large 3 675BMistral AI | 675B(41B active) | SS | 33.6 tok/s | 66.3 GB | |
minimax-m2.5MiniMax | 230B(10B active) | SS | 98.1 tok/s | 22.7 GB | |
| AdTools for AI BuildersAll the hardware and tools you need to run private AI at your office or at home.Shop Now | |||||
Gemma 4 12BGoogle | 12B | SS | 69.5 tok/s | 32.0 GB | |
Model Compatibility
Use the calculator to weigh this device against any other, model by model, with speed and fit scores.
Straight answers to the questions we hear most often.
Can't find what you're looking for? Book a Discovery Call
Hy3 is the highest-rated model that runs on Google Cloud TPU v5p at 4-bit, at about 40 tok/s. In total, Google Cloud TPU v5p can run 264 of the 298 open models we track.
We size each model at 4-bit (Q4_K_M), add the KV cache for its context length and a fixed runtime overhead, then compare that to this device VRAM. Each model is graded on how comfortably it fits and how fast it should run on this card.
It is an estimate of decode speed in tokens per second at 4-bit, based on this device memory bandwidth and each model size. Real speed depends on your runtime, batch size, and context length, so use it to compare models, not as a hard promise.
Lower-bit quantization shrinks a model so more of them fit. These grades already assume 4-bit, which is the common local default. Going below 4-bit can squeeze a larger model in at some quality cost; going above needs more VRAM and may not fit.
Mixture-of-experts models only activate a fraction of their parameters at a time, so their runtime memory is closer to the active size than the headline size. We score fit on the active parameters, which is why some very large models still fit.
