How much GPU memory does QwQ-R1984-32B need?

About 80.9 GB for conversations of up to 8,192 tokens: 65.53 GB of weights, the cache for one full-length conversation and the runtime's own memory (an estimate: its settings are behind its licence).

What is the cheapest GPU to run QwQ-R1984-32B?

Right now 4x RTX 3090 at $1.16 an hour; RTX PRO 6000 at $1.50 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.

Can QwQ-R1984-32B run on an RTX 4090?

Not on one: it needs about 80.9 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Models / QwQ-R1984

VIDraft/QwQ-R1984-32B

Text generation · served with vLLM · ↓ 1 downloads a month · View on Hugging Face

VIDraft/QwQ-R1984-32B needs about 80.9 GB of GPU memory for conversations of up to 8,192 tokens (an estimate: its settings are behind its licence). It runs on 4x RTX 3090 from $1.16/h right now; best value: RTX PRO 6000 at $1.50/h.

The model's publisher asks users to accept its licence (apache-2.0). GridPup cannot download this one yet: choose an ungated version below.

Weights: 66 GB
GPU memory needed: about 80.9 GB (estimate)
Conversation length: 8,192 tokens
Memory · disk: 40.8 GB · 113 GB
Architecture: Qwen2ForCausalLM
Licence: apache-2.0

Hardware and deploy

The publisher has not let GridPup download this licence-gated model yet. An ungated version of it (above) can be deployed.

Questions

How much GPU memory does QwQ-R1984-32B need?: About 80.9 GB for conversations of up to 8,192 tokens: 65.53 GB of weights, the cache for one full-length conversation and the runtime's own memory (an estimate: its settings are behind its licence).
What is the cheapest GPU to run QwQ-R1984-32B?: Right now 4x RTX 3090 at $1.16 an hour; RTX PRO 6000 at $1.50 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.
Can QwQ-R1984-32B run on an RTX 4090?: Not on one: it needs about 80.9 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Loading…