LocalLLMChecker

Desktop GPU · AMD

What AI models can a Radeon RX 7900 XT run?

20 GB is the enthusiast standard, and it shows: a 32B model fits at a good quantization with a long context, and a 70B is reachable if you are willing to drop precision. At 800 GB/s there is enough bandwidth to keep generation responsive for anything that fits.

Memory20 GB
Bandwidth800 GB/s
Usable for a model19.2 GB
Runtime backendROCM
ArchitectureRDNA 3
Power315 W

The short answer

Assuming an 8k context window and default settings, these are the models worth downloading first.

Best all-rounder

Qwen3.5 27B

Q5_K_M · 18.64 GB · about 27.4 tokens/s

Best for code

Qwen3 32B

IQ4_XS · 18.26 GB · about 28 tokens/s

Largest that still runs well

Qwen3.5 35B A3B

35.95B parameters · IQ4_XS · about 177.5 tokens/s

Fastest useful answer

Qwen3.5 35B A3B

about 177.5 tokens/s · 18.09 GB

See how fast it feels

Radeon RX 7900 XT running Qwen3.5 27B at Q5_K_M

Wait for the first word987 ms
Then writes at27.4 tok/s
Whole answer7.3 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Where it sits

With 20 GB it belongs to the 16 GB tier. The 16 GB guide explains what that tier runs and where it runs out, and the bandwidth guide explains why 800 GB/s decides how fast it answers.

Our indicative price for this card is $650 new, reviewed 2026-09-01. That is 3.1 GB of memory and 123 GB/s of bandwidth for every $100. Prices move, so treat these as a place to start a search and not a quote.

Devices in a similar price band, fastest memory first:

DevicePriceMemoryBandwidthBest all-round pick
GeForce RTX 3090 Ti $800 used 24 GB 1008 GB/s Qwen3.5 27B, 35.5 t/s
GeForce RTX 3090 $700 used 24 GB 936 GB/s Qwen3.5 27B, 33 t/s
GeForce RTX 5070 Ti $780 16 GB 896 GB/s Qwen3.5 27B, 47.9 t/s
GeForce RTX 5070 $560 12 GB 672 GB/s Gemma 4 12B Instruct, 47.3 t/s

Every model, scored on this device

Each row uses the highest-quality quantization that both fits and stays conversational. Speed is a single-stream estimate at 8k context.

Model Params Verdict Download Memory used Speed Max context
Qwen3.5 27B
Qwen 		27.78B Just fits Q5_K_M 18.64 GB 27.4 t/s 16k
Qwen3 32B
Qwen 		32.76B Just fits IQ4_XS 18.26 GB 28 t/s 8k
Granite 4.1 30B
Granite 		28.87B Just fits Q4_K_M 18.64 GB 27.4 t/s 8k
Gemma 4 31B Instruct
Gemma 		31.27B Runs great Q3_K_M 17.96 GB 28.5 t/s 8k
Gemma 4 26B A4B Instruct
Gemma 		25.81B (3.74B active) Just fits Q5_K_M 18.28 GB 87.8 t/s 16k
Qwen2.5 32B Instruct
Qwen 		32.76B Just fits IQ4_XS 18.26 GB 28 t/s 8k
Qwen2.5 Coder 32B Instruct
Qwen 		32.76B Just fits IQ4_XS 18.26 GB 28 t/s 8k
Gemma 4 12B Instruct
Gemma 		11.96B Runs great Q8_0 13.55 GB 38 t/s 32k
Gemma 3 27B Instruct
Gemma 		27.43B Runs great Q4_K_M 16.96 GB 30.2 t/s 32k
Mistral Small 24B Instruct 2501
Mistral 		23.57B Just fits Q6_K 19.18 GB 26.6 t/s 8k
Qwen3 14B
Qwen 		14.77B Runs great Q8_0 15.54 GB 33.1 t/s 16k
GPT OSS 20B
GPT-OSS 		20.91B (4.18B active) Runs great MXFP4 11.55 GB 121.2 t/s 128k
Qwen2.5 14B Instruct
Qwen 		14.77B Runs great Q8_0 15.79 GB 32.5 t/s 16k
Qwen2.5 Coder 14B Instruct
Qwen 		14.77B Runs great Q8_0 15.79 GB 32.5 t/s 16k
GLM 4.7 Flash
GLM 		31.22B (3.66B active) Runs great Q4_K_M 18.13 GB 118.2 t/s 16k
Qwen3.5 9B
Qwen 		9.65B Runs great Q8_0 9.25 GB 56.5 t/s 256k
Codestral 22B V0.1
Mistral 		22.25B Just fits Q6_K 19.06 GB 26.9 t/s 8k
Qwen3.5 35B A3B
Qwen 		35.95B (2.9B active) Runs great IQ4_XS 18.09 GB 177.5 t/s 64k
Phi 4
Phi 		14.66B Runs great Q8_0 16 GB 32.1 t/s 16k
Qwen3 30B A3B
Qwen 		30.53B (3.34B active) Runs great Q4_K_M 18.08 GB 111 t/s 16k
Gemma 3 12B Instruct
Gemma 		12.19B Runs great Q8_0 13.27 GB 38.8 t/s 64k
Qwen3 Coder 30B A3B Instruct
Qwen 		30.53B (3.34B active) Runs great Q4_K_M 18.08 GB 111 t/s 16k
Gemma 4 E4B Instruct
Gemma 		8B Runs great Q8_0 8.42 GB 61.9 t/s 128k
Granite 4.1 8B
Granite 		8.79B Runs great Q8_0 10.35 GB 50.2 t/s 64k
Qwen3 8B
Qwen 		8.19B Runs great Q8_0 9.02 GB 58 t/s 32k
Mistral Nemo Instruct 2407
Mistral 		12.25B Runs great Q8_0 13.15 GB 39.3 t/s 32k
Qwen2.5 VL 7B Instruct
Qwen 		8.29B Runs great Q8_0 8.49 GB 61.6 t/s 64k
Qwen2.5 7B Instruct
Qwen 		7.62B Runs great Q8_0 7.83 GB 67.1 t/s 32k
Llama 3.1 8B Instruct
Llama 		8.03B Runs great Q8_0 8.83 GB 59.2 t/s 64k
DeepSeek Coder V2 Lite Instruct
DeepSeek 		15.71B (2.74B active) Runs great Q8_0 15.91 GB 99.4 t/s 64k
Show the other 56 models
Gemma 2 9B Instruct 9.24B Runs great Q8_0 11.5 GB 44.9 t/s 8k
Qwen2.5 Coder 7B Instruct 7.62B Runs great Q8_0 7.83 GB 67.1 t/s 32k
OLMo 3 7B Instruct 7.3B Runs great Q8_0 9.72 GB 53.6 t/s 64k
Internlm3 8B Instruct 8.8B Runs great Q8_0 8.97 GB 58.3 t/s 32k
Mistral 7B Instruct V0.3 7.25B Runs great Q8_0 8.45 GB 62.1 t/s 32k
Gemma 4 E2B Instruct 5.12B Runs great Q8_0 5.44 GB 97.5 t/s 128k
Qwen3.5 4B 4.66B Runs great Q8_0 5.27 GB 101.7 t/s 256k
Mistral 7B Instruct V0.2 7.24B Runs great Q8_0 8.44 GB 62.1 t/s 32k
DeepSeek Coder 7B Instruct V1.5 6.91B Runs great Q8_0 10.58 GB 49.1 t/s 4k
DeepSeek Coder 6.7B Instruct 6.74B Runs great Q8_0 10.94 GB 47.4 t/s 16k
LFM2.5 8B A1B 8.47B Runs great Q8_0 8.81 GB 175.8 t/s 64k
Qwen3 4B 4.02B Runs great Q8_0 5.46 GB 97.8 t/s 32k
Gemma 3 4B Instruct 4.3B Runs great Q8_0 4.91 GB 109.7 t/s 128k
Phi 4 Mini Instruct 3.84B Runs great Q8_0 5.17 GB 104 t/s 64k
Granite 4.1 3B 3.4B Runs great Q8_0 4.34 GB 125.1 t/s 128k
Qwen2.5 3B Instruct 3.09B Runs great Q8_0 3.68 GB 149.5 t/s 32k
Phi 3 Mini 4k Instruct 3.82B Runs great Q8_0 4.8 GB 112.6 t/s 4k
Llama 3.2 3B Instruct 3.21B Runs great Q8_0 4.42 GB 123.2 t/s 128k
LFM2.5 2.6B 2.7B Runs great Q8_0 3.14 GB 178.5 t/s 128k
Qwen3.5 2B 2.27B Runs great Q8_0 2.7 GB 211.7 t/s 256k
Gemma 2 2B Instruct 2.61B Runs great Q8_0 3.54 GB 156.4 t/s 8k
Starcoder2 3B 3.03B Runs great Q8_0 3.34 GB 168.3 t/s 16k
Qwen3 1.7B 2.03B Runs great Q8_0 3.22 GB 173.1 t/s 32k
DeepSeek R1 Distill Qwen 1.5B 1.78B Runs great Q8_0 2.07 GB 285.4 t/s 128k
Qwen2.5 1.5B Instruct 1.54B Runs great Q8_0 2.06 GB 286.5 t/s 32k
SmolLM2 1.7B 1.71B Runs great Q8_0 3.53 GB 156.4 t/s 8k
OLMo 2 0425 1B 1.48B Runs great Q8_0 2.6 GB 220.8 t/s 4k
Llama 3.2 1B Instruct 1.24B Runs great Q8_0 1.82 GB 338 t/s 128k
LFM2.5 1.2B Instruct 1.17B Runs great Q8_0 1.46 GB 446.2 t/s 64k
Gemma 3 1B Instruct 1B Runs great Q8_0 1.35 GB 480.5 t/s 32k
Qwen3.5 0.8B 0.87B Runs great Q8_0 1.28 GB 513.1 t/s 256k
Qwen3 0.6B 0.75B Runs great Q8_0 1.92 GB 308.7 t/s 32k
Qwen2.5 0.5B Instruct 0.49B Runs great Q8_0 0.88 GB 862.7 t/s 32k
LFM2.5 350M 0.35B Runs great Q8_0 0.68 GB 1335.9 t/s 64k
Qwen3 Coder Next 79.67B Partial offload n/a 45.22 GB 46 t/s n/a
Qwen3 Next 80B A3B Instruct 81.32B Partial offload n/a 46.15 GB 44.9 t/s n/a
Phi 3.5 MoE Instruct 41.87B Partial offload n/a 24.87 GB 26.8 t/s n/a
GPT OSS 120B 116.83B Partial offload n/a 59.4 GB 21 t/s n/a
Qwen3.5 122B A10B 125.09B Partial offload n/a 70.61 GB 15.7 t/s n/a
Mixtral 8x7B Instruct V0.1 46.7B Partial offload n/a 27.59 GB 11 t/s n/a
GLM 4.5 Air 110.47B Partial offload n/a 56.19 GB 10.4 t/s n/a
MiniMax M2.5 228.7B Partial offload n/a 130.56 GB 7.6 t/s n/a
MiniMax M2.7 228.69B Partial offload n/a 130.55 GB 7.6 t/s n/a
MiniMax M3 427.04B Partial offload n/a 240.84 GB 3.9 t/s n/a
Qwen3 235B A22B 235.09B Partial offload n/a 133.71 GB 3.8 t/s n/a
GLM 4.5 358.34B Partial offload n/a 204.37 GB 2.8 t/s n/a
Kimi K2.6 1026.88B Partial offload n/a 577.78 GB 2.4 t/s n/a
Kimi K2 Instruct 1026.41B Partial offload n/a 577.52 GB 2.4 t/s n/a
DeepSeek V3.2 685.4B Partial offload n/a 385.91 GB 2.4 t/s n/a
DeepSeek R1 684.53B Partial offload n/a 385.42 GB 2.4 t/s n/a
GLM 5.1 753.86B Partial offload n/a 424.51 GB 2.2 t/s n/a
GLM 5.2 753.33B Partial offload n/a 424.21 GB 2.1 t/s n/a
Qwen2.5 VL 72B Instruct 73.41B Partial offload n/a 43.62 GB 1.2 t/s n/a
Qwen2.5 72B Instruct 72.71B Partial offload n/a 43.22 GB 1.2 t/s n/a
Llama 3.3 70B Instruct 70.55B Partial offload n/a 42.12 GB 1.2 t/s n/a
Llama 3 3 Nemotron Super 49B V1 49.87B Partial offload n/a 47.99 GB 1 t/s n/a

What it will not run

0 of the 86 architectures we track are out of reach here, even at 2-bit precision. Another 22 run by splitting the model between the card and system RAM, which works but drops generation to single digits for most dense models.

If you want more headroom

The next steps up in memory, in order. More memory changes which models load at all; more bandwidth changes how fast they answer, so the two are worth weighing separately.

DeviceMemoryBandwidthModels that fit
GeForce RTX 4090 24 GB 1008 GB/s 69 Check price
GeForce RTX 3090 Ti 24 GB 1008 GB/s 69 Check price
Radeon RX 7900 XTX 24 GB 960 GB/s 69 Check price
GeForce RTX 3090 24 GB 936 GB/s 69 Check price

Price links go to an Amazon search for the model name and are affiliate links: if you buy through one, we earn a commission at no cost to you. We do not take payment for placement, and the ordering above is by memory capacity alone.

Questions people ask about an RX 7900 XT

How much VRAM does a Radeon RX 7900 XT have?

The Radeon RX 7900 XT has 20 GB of VRAM, of which roughly 19.2 GB is available to a model once the operating system takes its share.

Can a Radeon RX 7900 XT run a 70B model?

Only partially. About 48 of 48 layers fit on the card and the rest run on the CPU, which brings generation down to roughly 46 tokens per second.

What is the best local model for a Radeon RX 7900 XT?

Qwen3.5 27B at Q5_K_M is the strongest all-round fit: 18.64 GB used, around 27.4 tokens per second at 8k context.