Ovis2
Open vision-language models from 1B to 34B built on structural visual-textual embedding alignment.
About
Ovis2 is AIDC-AI's multimodal LLM family released January 26, 2025, spanning 1B to 34B parameters by pairing aimv2 vision encoders with Qwen2.5 language backbones. Its signature is structural embedding alignment: visual features are mapped into the text embedding space so the language model reasons over images more or less the way it reasons over words. The release added video and multi-image training, multilingual OCR, and stronger chain-of-thought reasoning over Ovis1.6, with a 32K multimodal context window.
The provider publishes full VLMEvalKit benchmark tables: Ovis2-8B scores 84.9 on MMBench, 83.6 on MMBench-V1.1, 57.4 on MMMU, 86.6 on AI2D, 89.1 on OCRBench, and 68.0/71.6 on VideoMME, broadly ahead of Qwen2.5-VL-7B and InternVL2.5-8B-MPO on the same tables. Those are provider-run numbers, but the methodology and harness are public. Weights are Apache 2.0, and Ovis2.5 (2B/9B with native-resolution perception) followed in August 2025.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- MMBench: 84.9, MMBench-V1.1: 83.6, MMMU (val): 57.4, AI2D: 86.6, OCRBench: 89.1, RealWorldQA: 72.5 (Ovis2-8B, provider-run via VLMEvalKit)
- VideoMME: 68.0/71.6, MLVU: 76.4, MVBench: 68.15 (Ovis2-8B, provider-run)
- Comparisons vs Qwen2.5-VL-7B and InternVL2.5-8B-MPO on the same tables are provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: real AIDC-AI release Jan 26, 2025 across six sizes with Apache-2.0 weights; full benchmark tables published on the model card with a public harness (VLMEvalKit); Ovis2.5 follow-up in Aug 2025 shows maintenance. Open weights. Strengths: transparent eval tables, permissive license, distinctive alignment architecture. Docked for: (1) benchmark tables are provider-run with no independent replication found; (2) modest community footprint versus Qwen-VL/InternVL peers.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · AIDC-AI release 2025-01-26; 1B–34B sizes; Apache-2.0 per model card
- · Full VLMEvalKit benchmark tables on model card (Ovis2-8B: MMBench 84.9, MMMU 57.4)
- · Ovis2.5 (2B/9B) followed Aug 2025
- · Tables are provider-run; smaller community footprint than Qwen-VL/InternVL