Best local AI models for AMD RX 7900 XT

The largest models that fit fully

The 30 largest of the 166 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.

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.

How to read this

The AMD RX 7900 XT graphics card comes equipped with 20 GB of GDDR6 memory. This onboard memory size is the most critical factor when running local AI models. To achieve fast processing speeds, the entire active portion of the model should fit directly inside this video memory. If a model exceeds this limit, your system must transfer data to system RAM, which slows down performance.

The quantization column indicates the compression level applied to each model. Raw models are often too large for consumer hardware, so they are quantized to reduce their footprint. For example, Gemma 3 27B, Gemma 3 4B/12B/27B (vision), and Wan 2.2 / 2.5 can run at the Q4_K_M quantization level. This compression uses 19.8 GB of video memory, which fits just inside your hardware limit. Similarly, Gemma 4 26B-A4B and Gemma 4 (all sizes) fit well at Q4_K_M by using 19 GB of video memory.

As models get smaller, you can use higher quality quantizations. Aria at 25B fits at Q4_K_M using 18.3 GB. Mistral Small 3.2, Magistral Small, and Devstral Small 1.1 use 17.6 GB at Q4_K_M. You can run Solar Pro and Codestral 22B at the Q5_K_M level using 18.7 GB. Models like gpt-oss-20b and Reka Flash 3 use 17.9 GB at Q5_K_M. Qwen-Image and Qwen-Image-Edit run at Q6_K using 19.7 GB, while CogVLM2 uses 18.7 GB at Q6_K. HunyuanImage 2.1 / 3.0 uses 16.7 GB at Q6_K, and Ling-Coder-Lite uses 16.5 GB at Q6_K.

Highly compressed smaller models can run at the Q8_0 level for maximum precision. DeepSeek-Coder-V2 16B / 236B and Kimi-VL A3B run at Q6_K using 15.7 GB. Apriel-1.5-15B-Thinker and StarCoder2 3B / 7B / 15B run at Q8_0 using 19.1 GB. Qwen2.5 14B uses 19.5 GB at Q8_0. Phi-3 Medium, Phi-4, Phi-4-reasoning / -plus, Wan 2.2 T2I, Wan 2.1 (1.3B / 14B), and SkyReels V2 all use 17.8 GB at Q8_0. Vicuna 13B fits at Q8_0 using 16.5 GB.

When a model is too large for the 20 GB video memory, you must offload parts of it to your system RAM. This offloading allows you to run larger models but reduces your generation speed. For these cases, we assume you have 32 GB of system RAM. Qwen3-30B-A3B and Qwen3-Coder 30B-A3B need 22 GB at Q4_K_M and require 24 GB of system RAM. Qwen3 8B / 14B / 32B, Qwen3.5 (dense variants), Aya Expanse 8B / 32B, Granite 4.0 Small/Tiny, and Qwen2.5-Coder 0.5B to 32B need 23.4 GB at Q4_K_M and require 25.4 GB of system RAM. OTel 2.0 LLM 31B IT needs 27.5 GB at Q4_K_M and requires 29.5 GB of system RAM. DeepSeek-Coder 1.3B / 6.7B / 33B and WizardCoder 33B need 24.2 GB at Q4_K_M and require 26.2 GB of system RAM.

All memory calculations shown here are based on a standard 4k context window. If you increase the context window to process longer documents or chat histories, the memory requirement will grow. You must leave some safety headroom in your video memory to prevent out of memory errors during long conversations.