What AI can a Radeon RX 7900 XTX run?

Quick verdict

PC with Radeon RX 7900 XTX (24 GB) and 32 GB RAM can run 23 of the 31 models tracked here at a comfortable speed, and another 3 slowly. The largest chat or coding model that fits is Gemma 4 31B, which needs about 21.5 GB and generates roughly 22 words a second. The quickest is Gemma 4 E2B at about 190+ words a second, which feels faster than you can read.

What limits this machine

Video memory (VRAM) is the limit: 24 GB on this card. A model must fit in VRAM to run fast. Anything larger spills into system RAM (32 GB in this setup), which works but is slow, because system RAM moves data at roughly 70 GB/s while the card moves 960 GB/s. AMD cards run most of these tools through ROCm or Vulkan, so setup is a little less smooth than on NVIDIA, and the image and video times here include a 1.5 times penalty.

What it is good for

Chat and coding

Best chat pick: Gemma 4 31B at about 22 words a second. Best coding pick: Qwen3 Coder 30B at about 119.

Images

• Stable Diffusion XL: runs great, about about 16 s per image
• FLUX.2 klein 4B: runs great, about about 28 s per image
• Z-Image Turbo 6B: runs great, about about 40 s per image

Video

• LTX-2.3: slow, about about 6 min per 4 second clip
• Wan 2.2 5B: runs great, about about 11 min per 4 second clip

Voice and music

• Whisper large-v3 turbo: runs great, about about 42 s per 10 minutes of audio
• Kokoro TTS: runs great, about about 6 s per minute of speech
• Qwen3 TTS 1.7B: runs great, about about 30 s per minute of speech
• ACE-Step 1.5: runs great, about about 1 min per 1 minute song

Fastest models here

• Gemma 4 E2B: about 190+ words a second, faster than you can read
• Gemma 4 E4B: about 135 words a second, faster than you can read
• Qwen 3.5 4B: about 131 words a second, faster than you can read
• Qwen3 Coder 30B: about 119 words a second, faster than you can read
• gpt-oss 20B: about 98 words a second, faster than you can read

Memory headroom

Spare memory after loading the model decides how long a conversation or document can be before things slow down. With about 42 GB usable here:

• Gemma 4 31B needs 21.5 GB, leaving 20.5 GB spare, plenty for long documents.
• DeepSeek R1 32B needs 21.5 GB, leaving 20.5 GB spare, plenty for long documents.
• Qwen3 Coder 30B needs 20.4 GB, leaving 21.6 GB spare, plenty for long documents.
• Qwen 3.6 27B needs 19.9 GB, leaving 22.1 GB spare, plenty for long documents.
• Gemma 4 26B A4B needs 19.4 GB, leaving 22.6 GB spare, plenty for long documents.

What does not fit, and by how much

• Llama 3.3 70B needs about 45.6 GB and this machine tops out near 42 GB, so it is about 4 GB short.
• gpt-oss 120B needs about 68.8 GB and this machine tops out near 42 GB, so it is about 27 GB short.
• Qwen 3.5 122B A10B needs about 85.5 GB and this machine tops out near 42 GB, so it is about 44 GB short.
• DeepSeek V3 671B needs about 424.7 GB and this machine tops out near 42 GB, so it is about 383 GB short.

How it compares

• PC with Radeon RX 9070 XT (16 GB) and 32 GB RAM: 7 fewer models; drops Gemma 4 26B A4B, Gemma 4 31B, Qwen3 Coder 30B and more.
• PC with RTX 3090, RTX 4090 or RTX 5090 Laptop and 32 GB RAM: the same number of models.
• PC with RTX 5090 (32 GB) and 32 GB RAM: 3 more models; adds Qwen 3.5 35B A3B, Qwen 3.6 35B A3B, LTX-2.3.

Is an upgrade worth it?

The biggest single jump is Ryzen AI Max+ 395 (Strix Halo), 128 GB, which would unlock 6 more models, including Qwen 3.5 35B A3B, Qwen 3.5 122B A10B, gpt-oss 120B.

Questions