16 GB at 624 GB/s → 16 GB at 640 GB/s

Same 16 GB either way. Radeon RX 7800 XT to Radeon RX 9070 XT does not unlock a larger catalog model at 8K, it raises bandwidth. Qwen2.5 14B Instruct moves from 32–56 tok/s to 33–57 tok/s.

Hands installing a graphics card into a PC motherboard slot

Radeon RX 7800 XT and Radeon RX 9070 XT, same 8K context and recommended quant, 32 GB system RAM.

ModelRadeon RX 7800 XTRadeon RX 9070 XTSpeed beforeSpeed after
Llama 3.2 1B InstructFitsFits319–618 tok/s327–634 tok/s
Llama 3.2 3B InstructFitsFits116–247 tok/s119–254 tok/s
Llama 3.1 8B InstructFitsFits57–101 tok/s59–104 tok/s
Llama 3.3 70B InstructNoNo——
Qwen2.5 7B InstructFitsFits67–107 tok/s68–109 tok/s
Qwen2.5 14B InstructFitsFits32–56 tok/s33–57 tok/s
Qwen2.5 32B InstructOffloadOffload2–3 tok/s4–6 tok/s
Qwen2.5 72B InstructNoNo——
Qwen3 8BFitsFits55–99 tok/s56–102 tok/s
Qwen3 14BFitsFits33–55 tok/s34–57 tok/s
Qwen3 32BOffloadOffload2–3 tok/s4–6 tok/s
Qwen3 30B-A3BOffloadOffload33–48 tok/s53–78 tok/s
Qwen3.5 9BFitsFits57–88 tok/s59–90 tok/s
Qwen3.8 27BOffloadOffload11–16 tok/s15–21 tok/s
Qwen3.6 35B-A3BOffloadOffload21–31 tok/s38–55 tok/s
Gemma 3 4B InstructFitsFits123–201 tok/s127–206 tok/s
Gemma 3 12B InstructFitsFits42–68 tok/s43–70 tok/s
Gemma 3 27B InstructOffloadOffload8–11 tok/s11–16 tok/s
Gemma 4 E2BFitsFits106–161 tok/s109–165 tok/s
Gemma 4 E4BFitsFits65–100 tok/s67–103 tok/s
Gemma 4 12BFitsFits45–70 tok/s46–72 tok/s
Gemma 4 26B-A4BOffloadOffload68–99 tok/s92–134 tok/s
Gemma 4 31BOffloadOffload3–5 tok/s6–8 tok/s
gpt-oss-20bFitsFits131–206 tok/s134–211 tok/s
gpt-oss-120bNoNo——
Phi-4FitsFits32–55 tok/s33–57 tok/s
Phi-4-mini InstructFitsFits96–200 tok/s99–205 tok/s
Mistral Nemo 12B InstructFitsFits39–67 tok/s40–68 tok/s
Mistral Small 3.2 24B InstructTightTight22–35 tok/s22–36 tok/s
Devstral Small 2 24B InstructTightTight22–35 tok/s22–36 tok/s
DeepSeek-R1 Distill Qwen 14BFitsFits32–56 tok/s33–57 tok/s
DeepSeek-R1 Distill Qwen 32BOffloadOffload2–3 tok/s4–6 tok/s
SmolLM3 3BFitsFits136–261 tok/s140–267 tok/s

FAQ

What counts as an upgrade here?

A model newly reaching Fits, or a higher tokens-per-second range when VRAM does not change. Gaming FPS is a different question.

Are the two cards measured the same way?

Both use 8,192 tokens, the recommended quant, and 32 GB of assumed system RAM. Bandwidth is 624 GB/s versus 640 GB/s.

Should I buy this for local AI?

Buy VRAM first when a model you want is in Offload or No. Buy bandwidth when the model already fits and you want more tokens per second.