RX 7900 XTX

Muse Glimmer 30B is the best use of this card: 18.7 GB of the 24 GB available, an estimated 39 tokens per second, comfortable for chat, and 131,072 tokens of context still available.

How this is chosen, since no benchmark is quoted: models are compared by size, a mixture-of-experts model counting at the geometric mean of its total and active parameters (a rule of thumb, not a measurement). The pick is the biggest class that fits with memory to spare and answers at 25 tokens per second or better, and within that class the most recent general-purpose release; coding and reasoning specialists are listed above but not recommended by default.

Qwen3.6 35B-A3B (MoE) is the largest model the card will hold, at 22.4 GB, but it is the wrong daily driver: filling 93% of memory leaves room for only 25,600 tokens of conversation. More parameters are not worth a context window that runs out mid-document. It is a mixture-of-experts model reading only 3B parameters per token, so it is the faster of the two at an estimated 386 tokens per second; what it gives up is the depth of a dense model that reads all of its weights for every token.

If quality matters more than parameter count, Gemma 4 12B fits at Q8_0 in about 14.2 GB. Q8 is near-lossless; a smaller model at Q8 often beats a larger one squeezed into Q3.

Quantising the KV cache to Q8 roughly doubles what you can hold, at a quality cost most people never notice: Nanbeige4.3 33B on this card goes from 10,240 to 21,504 tokens. Capped at 128k tokens, or at the model's own native window where that is smaller, marked "model limit": free memory beyond that point buys nothing. Rows that hold far more context than their size suggests are hybrid, sliding-window or latent-attention models, which cache only a fraction of what a standard transformer does; each model page shows the working.

The first model out of reach is Flint 40B-A2.5B (MoE): about 24.8 GB at Q4_K_M and 8k context, against 24 GB of usable memory. Dropping to Q3_K_M would need about 20.2 GB, which fits, though Q3 loses enough quality that a smaller model at Q4 is usually the better trade. Offloading the remainder to system RAM works and is ten to fifty times slower; it is a way to see a model run, not a way to use one.

Worth it if what you run is llama.cpp-based chat and you are comfortable checking each tool for ROCm support. If you want everything to work first time, including image generation and fine-tuning, the CUDA ecosystem decides it for a 3090 or 4090.

Retail prices drift, and a figure written today would be wrong within a month, so this page does not carry one. Put the figure you are actually looking at, with your electricity rate and your real monthly hours, into the build vs rent calculator, which already knows this card draws 355 W under load. It returns the month owning becomes cheaper, or tells you it never does. Hours per month decides it far more often than the price does.

• RTX 3090: 24 GB · 936 GB/s · best fit Muse Glimmer 30B at ~38 tok/s. The value benchmark for local models, and it has been for years.
• RTX 4090: 24 GB · 1,008 GB/s · best fit Muse Glimmer 30B at ~41 tok/s. The card most local-model advice is implicitly written for.
• RTX 5090: 32 GB · 1,792 GB/s · best fit Muse Glimmer 30B at ~73 tok/s. The first consumer card whose memory bandwidth changes what is comfortable.

All 40 cards are compared side by side on the GPU index, and the same numbers from the model's point of view are on each model fit table. To check one specific pairing rather than read a table, use the can-I-run-it checker.