16 GB at 672 GB/s → 16 GB at 960 GB/s

Same 16 GB either way. GeForce RTX 4070 Ti Super to GeForce RTX 5080 does not unlock a larger catalog model at 8K, it raises bandwidth. Qwen2.5 14B Instruct moves from 35–60 tok/s to 50–85 tok/s.

Hands installing a graphics card into a PC motherboard slot

GeForce RTX 4070 Ti Super and GeForce RTX 5080, same 8K context and recommended quant, 32 GB system RAM.

ModelGeForce RTX 4070 Ti SuperGeForce RTX 5080Speed beforeSpeed after
Llama 3.2 1B InstructFitsFits343–666 tok/s491–951 tok/s
Llama 3.2 3B InstructFitsFits125–266 tok/s178–380 tok/s
Llama 3.1 8B InstructFitsFits62–109 tok/s88–156 tok/s
Llama 3.3 70B InstructNoNo——
Qwen2.5 7B InstructFitsFits72–115 tok/s102–164 tok/s
Qwen2.5 14B InstructFitsFits35–60 tok/s50–85 tok/s
Qwen2.5 32B InstructOffloadOffload2–3 tok/s4–6 tok/s
Qwen2.5 72B InstructNoNo——
Qwen3 8BFitsFits59–107 tok/s85–153 tok/s
Qwen3 14BFitsFits36–60 tok/s51–85 tok/s
Qwen3 32BOffloadOffload2–3 tok/s4–6 tok/s
Qwen3 30B-A3BOffloadOffload33–49 tok/s61–89 tok/s
Qwen3.5 9BFitsFits62–95 tok/s88–135 tok/s
Qwen3.8 27BOffloadOffload11–16 tok/s19–27 tok/s
Qwen3.6 35B-A3BOffloadOffload21–31 tok/s41–59 tok/s
Gemma 3 4B InstructFitsFits133–216 tok/s190–308 tok/s
Gemma 3 12B InstructFitsFits45–74 tok/s65–105 tok/s
Gemma 3 27B InstructOffloadOffload8–11 tok/s14–20 tok/s
Gemma 4 E2BFitsFits114–173 tok/s163–247 tok/s
Gemma 4 E4BFitsFits70–108 tok/s101–154 tok/s
Gemma 4 12BFitsFits49–75 tok/s70–108 tok/s
Gemma 4 26B-A4BOffloadOffload70–102 tok/s118–172 tok/s
Gemma 4 31BOffloadOffload3–5 tok/s6–9 tok/s
gpt-oss-20bFitsFits141–222 tok/s201–317 tok/s
gpt-oss-120bNoNo——
Phi-4FitsFits34–59 tok/s49–85 tok/s
Phi-4-mini InstructFitsFits104–216 tok/s148–308 tok/s
Mistral Nemo 12B InstructFitsFits42–72 tok/s60–103 tok/s
Mistral Small 3.2 24B InstructTightTight24–38 tok/s34–54 tok/s
Devstral Small 2 24B InstructTightTight24–38 tok/s34–54 tok/s
DeepSeek-R1 Distill Qwen 14BFitsFits35–60 tok/s50–85 tok/s
DeepSeek-R1 Distill Qwen 32BOffloadOffload2–3 tok/s4–6 tok/s
SmolLM3 3BFitsFits147–281 tok/s210–401 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 672 GB/s versus 960 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.