Best GPU for Local AI Fine-tuning

Written by Jakub Rusinowski · Last updated July 21, 2026

Ranked for fine-tuning a base model: Picking a base model to adapt on your own data with LoRA/QLoRA.

Best overall: AMD Radeon RX 7900 XT

20 GB VRAM at 800 GB/s. It runs 69 of the models that qualify for this workload; the strongest is Gemma 4 31B at an estimated 28.6 tokens/sec.

The picks

	GPU	VRAM	MSRP	Models that fit	Best model it runs	Est. speed
Best overall	AMD Radeon RX 7900 XT	20 GB	$899	69	Gemma 4 31B	~28.6 tok/s
Best value	Intel Arc B570	10 GB	$219	56	Qwen 3 14B	~27.8 tok/s
Budget pick	Intel Arc B570	10 GB	$219	56	Qwen 3 14B	~27.8 tok/s
Most memory	NVIDIA DGX Spark	128 GB	$4,699	92	Qwen 3.5 122B-A10B	~25.8 tok/s

GPU	Score	VRAM	MSRP	Models fit	Est. speed	Tok/s per watt	Cost per model
AMD Radeon RX 7900 XT	67.3	20 GB	$899	69	~28.6	0.09	$13
NVIDIA GeForce RTX 5090	56.7	32 GB	$1,999	79	~57.8	0.1	$25
Intel Arc B580	52.4	12 GB	$249	57	~33	0.17	$4
AMD Radeon RX 7900 XTX	51.1	24 GB	$999	79	~32.9	0.09	$13
AMD Radeon RX 7800 XT	51.1	16 GB	$499	64	~43.9	0.17	$8
NVIDIA GeForce RTX 4090	50.9	24 GB	$1,599	79	~34.5	0.08	$20
AMD Radeon RX 9070	50.9	16 GB	$549	64	~44.9	0.2	$9
NVIDIA GeForce RTX 5070	50.5	12 GB	$549	57	~46.9	0.19	$10
AMD Radeon RX 9070 XT	50.2	16 GB	$599	64	~49.7	0.23	$9
NVIDIA GeForce RTX 5060	49.3	8 GB	$299	46	~49.5	0.34	$7
Intel Arc B570	49.1	10 GB	$219	56	~27.8	0.19	$4
NVIDIA GeForce RTX 3090	48.8	24 GB	$1,499	79	~32.2	0.09	$19

GPUs are scored on four axes: capability (45%), speed (30%), value (20%) and efficiency (5%), each normalised across the whole ranking.
Capability means the intrinsic strength of the best model the card can hold — not how many models fit, and not how fast it streams a small one.
Speed is capped at 40 tok/s: past that, more throughput does not change how the model feels to use.
Apple Silicon is excluded from this ranking. Those entries price a whole computer and rate a chip's package power, so on price-per-capability and performance-per-watt they would beat every add-in card by construction. Apple hardware is covered on the macOS platform page instead.
Ranked as a *base model* choice, not an assistant choice: licence weight is near-maximum (0.9) because a restrictive licence blocks distributing what you train, and latency is nearly ignored (0.2) because training throughput, not decode speed, sets the cost.

The AMD Radeon RX 7900 XT — 20 GB of VRAM runs 69 qualifying models, the strongest being Gemma 4.

The Intel Arc B570 at $219, which runs 56 qualifying models.

8 GB is the entry point at which a model for this workload will run at all. More memory buys a stronger model, not just a faster one.