Model reference · open weights
ConsiSpace is an open-weight language model from BAAI. ConsiSpace (BF16) weighs 11.4 GB; the smallest configuration that runs it is RTX 4060 Ti 16 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the BAAI/ConsiSpace model card, 2026-10-01
What it is
| Released by | BAAI |
|---|---|
| Released | 2026-09-04 |
| Parameters | 5.7B |
| VRAM | 11.4 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB | 11.6 GB | |||
| RTX 4060 Ti 16 GB | 6 | 1 | 53K | 15.4 GB |
| RTX 3090 24 GB | 33 | 8 | all 125K | 23.4 GB |
| RTX 4090 24 GB | 32 | 8 | all 125K | 23.4 GB |
| RTX 5090 32 GB | 58 | 14 | all 125K | 31.0 GB |
| L40S 48 GB | 101 | 25 | all 125K | 44.0 GB |
| A100 80 GB | 214 | 53 | all 125K | 78.2 GB |
| H100 80 GB | 200 | 50 | all 125K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 252 | 63 | all 125K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 297 | 74 | all 125K | 107 GB |
| H200 141 GB | 398 | 99 | all 125K | 138 GB |
| B200 180 GB | 523 | 130 | all 125K | 176 GB |
| 2× RTX 3060 12 GB tensor parallel | 25 | 6 | all 125K | 11.6 GB a card |
| 2× RTX 4060 Ti 16 GB tensor parallel | 51 | 12 | all 125K | 15.4 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 13.7 GB | 14.6 GB |
| 5 | 14.9 GB | 19.4 GB |
| 8 | 15.8 GB | 23.1 GB |
| 16 | 18.2 GB | 32.7 GB |
| 32 | 23.1 GB | 52.1 GB |
| 64 | 32.7 GB | 90.7 GB |
One card, with vLLM's small-card settings.
From the model card
Official model weights for ConsiSpace.
ConsiSpace is a geometry-consistent multimodal framework for spatial reasoning in long-form visual observations.
📄 Paper: https://arxiv.org/abs/2607.17599
@article{huang2026consispace,
title={ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning},
author={Ting Huang and Zhenyu Zhang and Wenyuan Huang and Jian Yang and Hao Tang},
year={2026},
eprint={2607.17599},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.17599},
}
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.