Best local AI models for AMD Pro Duo
4 GB HBM. At a 4k context, 81 of the 233 models in our catalog with verified parameter counts fit fully, up to Lumina-Next / Lumina-Image 2.0 at 5B parameters.
Check your own machine against every model →The largest models that fit fully
The 30 largest of the 81 models that fit; every smaller model in the catalog fits too. Best quant means the highest quality compression whose weights and 4k context both sit inside the memory.
| Model | Parameters | Best quant that fits | Memory used at 4k |
|---|---|---|---|
| Lumina-Next / Lumina-Image 2.0 | 5B | Q4_K_M | 3.7 GB |
| CogVideoX 2B / 5B | 5B | Q4_K_M | 3.7 GB |
| DeepSeek-VL2 | 4.5B | Q5_K_M | 3.8 GB |
| DeepFloyd IF | 4.3B | Q5_K_M | 3.7 GB |
| Phi-3.5-vision | 4.2B | Q5_K_M | 3.6 GB |
| Qwen3 4B | 4B | Q6_K | 3.9 GB |
| Gemma 3 4B | 4B | Q6_K | 3.9 GB |
| Gemma 4 E4B | 4B | Q6_K | 3.9 GB |
| MiniCPM 3 4B | 4B | Q6_K | 3.9 GB |
| Danube 3 4B | 4B | Q6_K | 3.9 GB |
| Fish Speech 1.5 / OpenAudio S1 | 4B | Q6_K | 3.9 GB |
| Phi-4-mini-instruct | 3.8B | Q6_K | 3.7 GB |
| Phi-3.5 Mini | 3.8B | Q6_K | 3.7 GB |
| OmniGen / OmniGen2 | 3.8B | Q6_K | 3.7 GB |
| SD Cascade (Würstchen v3) | 3.6B | Q6_K | 3.5 GB |
| SDXL Turbo | 3.5B | Q6_K | 3.4 GB |
| SDXL Lightning | 3.5B | Q6_K | 3.4 GB |
| ACE-Step | 3.5B | Q6_K | 3.4 GB |
| MusicGen small/medium/large | 3.3B | Q6_K | 3.2 GB |
| SmolLM3 3B | 3B | Q8_0 | 3.8 GB |
| Replit Code v1.5 3B | 3B | Q8_0 | 3.8 GB |
| Kandinsky 3.1 | 3B | Q8_0 | 3.8 GB |
| Voxtral Mini / Small | 3B | Q8_0 | 3.8 GB |
| Orpheus TTS | 3B | Q8_0 | 3.8 GB |
| Higgs Audio v2 | 3B | Q8_0 | 3.8 GB |
| Allegro | 2.8B | Q8_0 | 3.6 GB |
| Open-Sora Plan | 2.7B | Q8_0 | 3.4 GB |
| LFM2 1.2B / 2.6B | 2.6B | Q8_0 | 3.3 GB |
| Playground v2.5 | 2.6B | Q8_0 | 3.3 GB |
| Stable Diffusion 3.5 Medium | 2.5B | Q8_0 | 3.2 GB |
Close, but only with CPU offload
These need more than the card holds at their smallest practical quant, so part of the model runs from system memory (figures assume 32 GB of it). They work, several times slower.
| Model | Parameters | Memory at FP8 / optimized | System RAM at 4k |
|---|---|---|---|
| Stable Diffusion XL | 3.417B | 4.1 GB needed | 6.1 GB |
| Phi-3 Mini | 3.8B | 4.4 GB needed | 6.4 GB |
| Phi-4-multimodal | 5.6B | 4.1 GB needed | 6.1 GB |
| Magicoder-S-DS 6.7B | 6.7B | 4.9 GB needed | 6.9 GB |
| Mistral 7B | 7B | 5.7 GB needed | 7.7 GB |
| Qwen2.5 0.5B / 1.5B / 3B / 7B | 7B | 5.1 GB needed | 7.1 GB |
| OLMo 2 1B / 7B | 7B | 5.1 GB needed | 7.1 GB |
| Falcon 3 1B / 3B / 7B | 7B | 5.1 GB needed | 7.1 GB |
| Command R7B | 7B | 5.1 GB needed | 7.1 GB |
| OpenHermes 2.5 | 7B | 5.1 GB needed | 7.1 GB |
How to read this
The AMD Pro Duo graphics card features 4 GB of HBM. This memory size determines which local AI models you can run directly on the hardware. To fit models within this limit, you must use quantized versions. Quantization reduces the precision of model weights to save space. The best quant column shows the highest quality quantization level that fits safely inside the 4 GB boundary.
For fully local execution on the GPU, you can run models up to 5B parameters. Lumina-Next or Lumina-Image 2.0 and CogVideoX 2B or 5B both run at the 5B size using the Q4_K_M quant which uses 3.7 GB of memory. DeepSeek-VL2 at 4.5B parameters fits using the Q5_K_M quant with 3.8 GB used. DeepFloyd IF at 4.3B parameters uses 3.7 GB under the Q5_K_M quant. Phi-3.5-vision at 4.2B parameters uses 3.6 GB under the Q5_K_M quant.
Several 4B parameter models run efficiently using the Q6_K quant which uses 3.9 GB of memory. These models include Qwen3 4B, Gemma 3 4B, Gemma 4 E4B, MiniCPM 3 4B, Danube 3 4B, and Fish Speech 1.5 or OpenAudio S1. Phi-4-mini-instruct, Phi-3.5 Mini, and OmniGen or OmniGen2 at 3.8B parameters use 3.7 GB with the Q6_K quant. Image generation models like SD Cascade (Würstchen v3) at 3.6B parameters use 3.5 GB, while SDXL Turbo, SDXL Lightning, and ACE-Step at 3.5B parameters use 3.4 GB with the Q6_K quant. MusicGen small/medium/large at 3.3B parameters uses 3.2 GB with the Q6_K quant.
Models at the 3B parameter scale can use the higher quality Q8_0 quant which uses 3.8 GB of memory. This group includes SmolLM3 3B, Replit Code v1.5 3B, Kandinsky 3.1, Voxtral Mini or Small, Orpheus TTS, and Higgs Audio v2. For slightly smaller models, Allegro at 2.8B parameters uses 3.6 GB, Open-Sora Plan at 2.7B parameters uses 3.4 GB, LFM2 1.2B or 2.6B uses 3.3 GB, Playground v2.5 uses 3.3 GB, and Stable Diffusion 3.5 Medium at 2.5B parameters uses 3.2 GB. All of these use the Q8_0 quant.
When a model exceeds the 4 GB HBM limit, you must use CPU offloading. This process splits the model between your GPU and your system RAM. Offloading allows you to run larger models but it reduces processing speed because system RAM is slower than HBM. For these setups, we assume your computer has 32 GB of system RAM available.
Under CPU offload, Stable Diffusion XL at 3.417B needs 4.1 GB at FP8 or optimized settings and requires 6.1 GB of system RAM. Phi-3 Mini at 3.8B needs 4.4 GB at Q4_K_M and requires 6.4 GB of system RAM. Phi-4-multimodal at 5.6B needs 4.1 GB at Q4_K_M and requires 6.1 GB of system RAM. Magicoder-S-DS 6.7B needs 4.9 GB at Q4_K_M and requires 6.9 GB of system RAM. Mistral 7B needs 5.7 GB at Q4_K_M and requires 7.7 GB of system RAM.
Other 7B models also run via offload using the Q4_K_M quant. Qwen2.5 0.5B or 1.5B or 3B or 7B, OLMo 2 1B or 7B, Falcon 3 1B or 3B or 7B, Command R7B, and OpenHermes 2.5 all need 5.1 GB of memory and require 7.1 GB of system RAM. Note that these memory calculations are based on a standard 4k context window. Running longer context windows will increase memory usage and may exceed your limits.