Model reference · open weights
Prompt-Refine is an open-weight language model from HiDream-ai. Prompt-Refine (BF16) weighs 62.5 GB; the smallest configuration that runs it is A100 80 GB.
What it is
| Released by | HiDream-ai |
|---|---|
| Released | 2026-05-13 |
| Parameters | 31.3B |
| VRAM | 62.5 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB … L40S 48 GB 6 smaller cards | — | — | — | |
| A100 80 GB | 2 | 1 | 54K | 78.2 GB |
| H100 80 GB | — | — | — | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 2 | 1 | 53K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 4 | 1 | 133K | 107 GB |
| H200 141 GB | 8 | 3 | all 256K | 138 GB |
| B200 180 GB | 12 | 3 | all 256K | 176 GB |
| 4× RTX 4090 24 GB tensor parallel | 3 | 2 | 105K | 23.4 GB a card |
| 4× RTX 3090 24 GB tensor parallel | 3 | 2 | 106K | 23.4 GB a card |
| 4× RTX 5090 32 GB tensor parallel | 9 | 5 | all 256K | 31.0 GB a card |
| 2× H100 80 GB tensor parallel | 9 | 3 | all 256K | 78.1 GB a card |
| 2× A100 80 GB tensor parallel | 16 | 9 | all 256K | 78.2 GB a card |
| 4× L40S 48 GB tensor parallel | 18 | 10 | all 256K | 44.0 GB a card |
| 2× RTX PRO 6000 Blackwell 96 GB tensor parallel | 13 | 5 | all 256K | 93.8 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 70.3 GB | 74.4 GB |
| 5 | 92.5 GB | 113 GB |
| 8 | 109 GB | 141 GB |
| 16 | 153 GB | 218 GB |
| 32 | 242 GB | 371 GB |
| 64 | 419 GB | 677 GB |
One card, with vLLM's small-card settings.
From the model card
Prompt Refine model for HiDream-O1-Image-Dev-2604. Finetuned from gemma-4-31b-it
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.