- Services
Every open model in our directory, ranked by how well it runs on AMD Instinct MI350P (144GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
AMD Instinct MI350P has 144 GB of VRAM, enough to run 321 of the 381 open models we track at 4-bit. The largest that still fits well is Falcon 180B (180B), and the top-rated is Qwen3.5-122B-A10B at about 118 tok/s.
321 of 381 models run comfortably on AMD Instinct MI350P.
See full AMD Instinct MI350P specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| Qwen3.5-122B-A10BAlibaba | SS | 118.1 tok/s | 92.2 GB |
| Mixtral 8x22B InstructMistral AI | SS | 73.9 tok/s | 102.7 GB |
| KolibriAleph Alpha | SS | 122.7 tok/s | 69.7 GB |
| Carnice-V2-27bkai-os | SS | 44.2 tok/s | 72.8 GB |
| Qwen3.6-27BAlibaba | SS | 44.2 tok/s | 72.8 GB |
| Qwen3.5-27BAlibaba | SS | 44.2 tok/s | 72.8 GB |
| AREX-2BAAI | SS | 43.7 tok/s | 73.7 GB |
| Gemma 4 31B ITGoogle | SS | 39.3 tok/s | 82.0 GB |
| AliceAI-Foundation-80B-A3B-Baseyandex | SS | 377.4 tok/s | 53.9 GB |
| Llama 3 70B InstructMeta | SS | 70.5 tok/s | 45.7 GB |
| Qwen3-32BAlibaba | SS | 59.7 tok/s | 53.9 GB |
| Nemotron 3.5 Bolt 74B-A9B 24GBNVIDIA | SS | 535.3 tok/s | 43.7 GB |
Qwen3.5-122B-A10BAlibaba | 122B(10B active) | SS | 118.1 tok/s | 92.2 GB | |
Mixtral 8x22B InstructMistral AI | 141B(39B active) | SS | 73.9 tok/s | 102.7 GB | |
KolibriAleph Alpha | 78B(3B active) | SS | 122.7 tok/s | 69.7 GB | |
Carnice-V2-27bkai-os | 27B | SS | 44.2 tok/s | 72.8 GB | |
Qwen3.6-27BAlibaba | 27B | SS | 44.2 tok/s | 72.8 GB | |
Qwen3.5-27BAlibaba | 27B | SS | 44.2 tok/s | 72.8 GB | |
AREX-2BAAI | 27.36B | SS | 43.7 tok/s | 73.7 GB | |
Gemma 4 31B ITGoogle | 31B | SS | 39.3 tok/s | 82.0 GB | |
Ad Tools for AI BuildersAll the hardware and tools you need to run private AI at your office or at home.Shop Now | |||||
| 81.29B(3B active) | SS | 377.4 tok/s | 53.9 GB | ||
| 70B | SS | 70.5 tok/s | 45.7 GB | ||
Qwen3-32BAlibaba | 32.8B | SS | 59.7 tok/s | 53.9 GB | |
| 74B(9B active) | SS | 535.3 tok/s | 43.7 GB | ||
| 74B(9B active) | SS | 535.3 tok/s | 43.7 GB | ||
Apertus 70BEPFL, ETH Zurich, CSCS | 70B | SS | 74.2 tok/s | 43.4 GB | |
Llama 2 70B ChatMeta | 70B | SS | 74.2 tok/s | 43.4 GB | |
| 29.99B | SS | 65.3 tok/s | 49.3 GB | ||
Ad Vast.aiAffordable on-demand GPU rentals for training and inference. Pick from thousands of hosts.Rent a GPU | |||||
Muse Glimmer-30BMeta | 30B | SS | 65.2 tok/s | 49.4 GB | |
Comet 60BMultiverse Computing | 61.1B(12.8B active) | SS | 385.9 tok/s | 36.4 GB | |
GoTo-AI 27B InstructGoTo Group | 27B | SS | 73.5 tok/s | 43.8 GB | |
Gemma 3 27B ITGoogle | 27B | SS | 73.5 tok/s | 43.8 GB | |
Mixtral 8x7B InstructMistral AI | 46.7B(12.9B active) | SS | 283.4 tok/s | 31.0 GB | |
Mistral Small 3 24BMistral AI | 24B | SS | 82.6 tok/s | 39.0 GB | |
K2-Horizon-MoVA-36B-A4BInstitute of Foundation Models | 36B(4B active) | SS | 164.3 tok/s | 38.2 GB | |
LLaMA 65BMeta | 65B | SS | 82.0 tok/s | 39.3 GB | |
| AdTools for AI BuildersAll the hardware and tools you need to run private AI at your office or at home.Shop Now | |||||
Recursive FlintRecursive | 39.8B(2.5B active) | SS | 447.7 tok/s | 28.8 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
Qwen3.5-122B-A10B is the highest-rated model that runs on AMD Instinct MI350P at 4-bit, at about 118 tok/s. In total, AMD Instinct MI350P can run 321 of the 381 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.
