Hardware → models
RTX 5090 32 GB
Start with gpt-oss-20b at MXFP4 profile for general chat. For a coding-focused option, try Qwen3-Coder 30B-A3B Instruct.
Use a current NVIDIA driver and a runtime build with Blackwell support. Check GPU detection and offloaded layers before relying on this fit.
Models to start with
Planning estimates32 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 | 16.2 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 | 10.2 GB |
| Qwen3 32BQ4_K_M | Larger general-purpose modelMore memory than 14B. Keep thinking mode fixed when comparing answers. | 21.6 GB | 9.4 GB |
| Llama 3.2 3B InstructQ4_K_M | Small text assistantA smaller weight footprint. Check the exact Llama checkpoint and licence. | 3.7 GB | 27.3 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.
Use a current NVIDIA driver and a runtime build with Blackwell support. Check GPU detection and offloaded layers before relying on this fit. CPU offloading changes the memory allocation and may slow generation.
Setup and common questions
Which RTX 5090 does this cover?
The desktop RTX 5090 has 32 GB GDDR7 and uses Blackwell. This page does not cover the laptop GPU. Compare the dedicated memory on your exact card before following the shortlist.
Why might a model fit but fail to use the GPU?
Use an NVIDIA driver and a runtime build that support Blackwell. Older CUDA builds can lack the required GPU support. Check that the runtime detects your card and reports layers loaded onto it; memory fit alone cannot establish compatibility.
What does 32 GB change?
It leaves more space around the 30–32B Q4 planning profiles than a 24 GB card. Compare Q5 or longer context in the Finder before increasing either. This site has no reviewed RTX 5090 timing measurements to establish how fast those settings run.
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.