Dolphin
ByteDance's analyze-then-parse document model — 0.3B for the original, 3B for v2 — that turns pages into structured JSON and Markdown with parallel element decoding.
About
Dolphin takes a two-stage approach to document parsing: analyze the page first, then parse it, using heterogeneous anchor prompting so different element types (text, tables, formulas, code) each get a decoding strategy that suits them. Element-level parsing runs in parallel, which keeps the lightweight architecture fast, and page-level parsing emits structured JSON plus Markdown in natural reading order.
The line has moved quickly: the original 0.3B Dolphin (ACL 2025) was followed by Dolphin-1.5 in October 2025 with large accuracy gains at the same size, then Dolphin-v2 in December 2025, a 3B model with 21-element detection, attribute extraction, and dedicated formula and code parsing. It handles multi-page PDFs and photographed documents, not just clean digital pages.
One honest wrinkle: the Hugging Face model card declares MIT for the weights while the GitHub repo's LICENSE file carries Qwen Research License text, so the licensing is ambiguous and worth checking before commercial use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Provider-reported, OmniDocBench v1.5 (repo README): Dolphin-v2 (3B) overall 89.78, FormulaCDM 87.63, TableTEDS 87.02, TEDS-S 90.48; Dolphin-1.5 (0.3B) overall 85.06; original Dolphin (0.3B) overall 74.67
- Paper accepted at ACL 2025 ('Document Image Parsing via Heterogeneous Anchor Prompting')
- No independent replication published yet — all headline numbers are provider-reported
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 (TIER verified, metrics-v1). Evidence: ACL 2025 paper with a public repo at github.com/bytedance/Dolphin, weights on Hugging Face (ByteDance/Dolphin-v2), and three shipped versions in seven months (1.0, 1.5, v2) showing real maintenance. Strengths: peer-reviewed method, open weights and code, fast parallel parsing, strong provider-reported OmniDocBench numbers. Docked for: (1) all benchmark numbers are provider-reported; (2) license ambiguity — HF card says MIT while the repo LICENSE file text is the Qwen Research License. License: ambiguous, verify before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ACL 2025 paper ('Document Image Parsing via Heterogeneous Anchor Prompting') from ByteDance; public repo github.com/bytedance/Dolphin and weights on HF (ByteDance/Dolphin-v2).
- · Three shipped versions in seven months (1.0 -> 1.5 -> v2, Dec 2025, 3B with 21-element detection); supports vLLM and TensorRT-LLM.
- · Provider-reported OmniDocBench v1.5: v2 overall 89.78, TableTEDS 87.02. No independent replication; license ambiguity docked.