Model reference · open weights
CASA-Qwen2_5-VL-Shared is an open-weight language model from kyutai. CASA-Qwen2_5-VL-3B-Shared (BF16) weighs 7.5 GB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the kyutai/CASA-Qwen2_5-VL-3B-Shared model card, 2026-10-01
What it is
| Released by | kyutai |
|---|---|
| Released | 2026-08-26 |
| Parameters | 3.8B |
| VRAM | 7.5 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB | 6 | 1 | 55K | 11.6 GB |
| RTX 4060 Ti 16 GB | 19 | 4 | all 125K | 15.4 GB |
| RTX 3090 24 GB | 46 | 11 | all 125K | 23.4 GB |
| RTX 4090 24 GB | 45 | 11 | all 125K | 23.4 GB |
| RTX 5090 32 GB | 71 | 17 | all 125K | 31.0 GB |
| L40S 48 GB | 114 | 28 | all 125K | 44.0 GB |
| A100 80 GB | 227 | 56 | all 125K | 78.2 GB |
| H100 80 GB | 213 | 53 | all 125K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 265 | 66 | all 125K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 310 | 77 | all 125K | 107 GB |
| H200 141 GB | 411 | 102 | all 125K | 138 GB |
| B200 180 GB | 536 | 134 | all 125K | 176 GB |
| 2× RTX 3060 12 GB tensor parallel | 38 | 9 | all 125K | 11.6 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 9.8 GB | 10.7 GB |
| 5 | 11.0 GB | 15.5 GB |
| 8 | 11.9 GB | 19.2 GB |
| 16 | 14.3 GB | 28.8 GB |
| 32 | 19.2 GB | 48.2 GB |
| 64 | 28.8 GB | 86.8 GB |
One card, with vLLM's small-card settings.
From the model card
This repository contains the model weights for CASA-Qwen2_5-VL-3B-Shared, introduced in the paper CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion.
This is a variant of CASA-Qwen2_5-VL-3B where the self-attention and cross-attention layers share the same parameters.
See CASA-Qwen2_5-VL-3B for more information
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.