Ovis models
6 Ovis models from ATH-MaaS on Hugging Face, from 853M to 10.2B parameters, published by ATH-MaaS in BF16, with contexts from 8K to 256K tokens. The smallest official model, OvisOCR2, needs about 1.9 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Pick a size
One row per official Ovis size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| OvisOCR2 | 853M | 1.9 GB | 1.0 GB | 0.5 GB | RTX 3070 | $0.088/hr | 0.38 GB |
| Ovis2-1B | 1.3B | 2.8 GB | 1.4 GB | 0.7 GB | RTX 3070 | $0.088/hr | not published for this architecture |
| Ovis1.6-Llama3.2-3B | 4.1B | 9.3 GB | 4.6 GB | 2.3 GB | RTX 3060 | $0.110/hr | not published for this architecture |
| Ovis2-4B | 4.6B | 10.3 GB | 5.2 GB | 2.6 GB | RTX 3060 | $0.110/hr | not published for this architecture |
| Ovis2.5-9B | 9.2B | 20.5 GB | 10.3 GB | 5.1 GB | RTX A5000 | $0.176/hr | not published for this architecture |
| Ovis1.6-Gemma2-9B | 10.2B | 22.8 GB | 11.4 GB | 5.7 GB | RTX A5000 | $0.176/hr | not published for this architecture |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (6)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| OvisOCR2 | 853M | BF16 | 1.9 GB | RTX 3070 | $0.088/hr |
| Ovis2-1B | 1.3B | BF16 | 2.8 GB | RTX 3070 | $0.088/hr |
| Ovis1.6-Llama3.2-3B | 4.1B | BF16 | 9.3 GB | RTX 3060 | $0.110/hr |
| Ovis2-4B | 4.6B | BF16 | 10.3 GB | RTX 3060 | $0.110/hr |
| Ovis2.5-9B | 9.2B | BF16 | 20.5 GB | RTX A5000 | $0.176/hr |
| Ovis1.6-Gemma2-9B | 10.2B | BF16 | 22.8 GB | RTX A5000 | $0.176/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.