Hardware → models
RX 7900 XTX 24 GB
Start with gpt-oss-20b at MXFP4 profile for general chat. For a coding-focused option, try Qwen3-Coder 30B-A3B Instruct.
Check your AMD backend and exact operating-system support first. CUDA builds are not an AMD GPU route; memory fit does not establish runtime compatibility.
Models to start with
Planning estimates24 GB dedicated VRAM · 4,000 tokens · 32 GB system RAM · full GPU residency
| Model / starting file | Why consider it | GPU memory estimate | Memory left¹ |
|---|---|---|---|
| gpt-oss-20bMXFP4 profile | Reasoning and tool-oriented tasksNeeds a runtime that supports MXFP4 and the Harmony conversation format. | 14.8 GB | 8.5 GB |
| Qwen3-Coder 30B-A3B InstructQ4_K_M | Coding mixture-of-experts modelFull weights must fit despite 3.3B active parameters. Use a compatible runtime. | 20.8 GB | 2.5 GB |
| Qwen3 32BQ4_K_M | Larger general-purpose modelMore memory than 14B. Keep thinking mode fixed when comparing answers. | 21.6 GB | 1.7 GB |
| Llama 3.2 3B InstructQ4_K_M | Small text assistantA smaller weight footprint. Check the exact Llama checkpoint and licence. | 3.7 GB | 19.6 GB |
From the published model guides here; the Finder covers the full catalogue. ¹ Memory left after the Finder’s reserve. Other applications can use this space. Larger models and higher precision do not guarantee better answers.
Check your AMD backend and exact operating-system support first. CUDA builds are not an AMD GPU route; memory fit does not establish runtime compatibility. CPU offloading changes the memory allocation and may slow generation.
Setup and common questions
Which AMD card is this?
The RX 7900 XTX has 24 GB GDDR6. The RX 7900 XT is a different card. These comparisons use dedicated VRAM; they do not add system RAM to the full-GPU budget.
Which backend should I choose?
Use a runtime with an AMD backend, such as llama.cpp HIP or Vulkan. For ROCm, check AMD’s matrix for your exact GPU, operating system and software version. A Linux support entry does not establish the same support on Windows or WSL.
How do I know inference is on the GPU?
Check the runtime log for GPU detection and offloaded layers. A model can load through CPU fallback without using the expected backend. Record that backend when sharing a test; results from a different GPU or software stack do not establish this card’s speed.
How are these recommendations chosen?
We use the main Finder’s balanced task selection and memory calculation, limited to models with published guides here. Task scores and speed proxies are planning inputs, not measured quality rankings.
GPU memory includes weights, context cache and a 0.9 GB runtime allowance, plus a 3% reserve (0.15–1.5 GB). Cache format and runtime buffers can change the fit. See the calculation assumptions.
Sources and related guides
Specification sources checked 3 October 2026.
These comparisons are planning estimates. No reviewed or independently reproduced measurements are available yet. About test evidence.