Florence-2
Tiny MIT-licensed vision-language model for dense vision tasks.
About
Florence-2 is Microsoft's lightweight vision foundation model from June 2024, in 0.23B and 0.77B sizes, trained on the FLD-5B dataset of 126M images with 5.4B annotations. It frames every vision task — captioning, detection, OCR, segmentation, phrase grounding — as prompt-based text generation, and punches well above its weight against much larger VLMs.
It is MIT licensed and genuinely open, which is rare at this quality level. The trade-off is age: newer VLMs have passed it on raw accuracy, and it predates the current agentic-vision wave, so expect to fine-tune it for niche domains.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Results comparable to much larger models such as Kosmos-2 on vision-language tasks (community reporting of Microsoft's figures)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Microsoft release, June 2024; MIT licensed — one of the few genuinely OSI-open vision foundation models at this quality; Widely used in document-understanding and grounding pipelines; still referenced as a lightweight baseline in 2025-2026; Model and training code open; FLD-5B dataset not released; Age is the main weakness: newer VLMs beat it on raw accuracy, but its license keeps it relevant. License: MIT — genuinely open, commercial use allowed.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Microsoft release, June 2024; MIT licensed — one of the few genuinely OSI-open vision foundation models at this quality
- · Widely used in document-understanding and grounding pipelines; still referenced as a lightweight baseline in 2025-2026
- · Model and training code open; FLD-5B dataset not released
- · Age is the main weakness: newer VLMs beat it on raw accuracy, but its license keeps it relevant