Datalab’s Lift is making waves as a 9-billion-parameter vision model built solely for structured JSON extraction from PDFs and images. It’s not an OCR tool, a PDF-to-Markdown converter, or a full enterprise document review platform. Lift’s focus is clear: schema-first document extraction.
Lift supports schema-constrained decoding. That means it returns JSON that matches the user’s exact schema — no guesswork, no extra fluff. This laser focus sets it apart from competitors that try to do more but deliver less precision.
NuMind’s NuExtract3 takes a different approach. At 4 billion parameters, it’s a smaller, vision-language reasoning model that blends structured extraction with image-to-Markdown conversion. It’s versatile but trades off accuracy for flexibility. In Datalab’s benchmark, Lift achieved 90.2% field accuracy. NuExtract3 clocked in at 81.5% — a notable gap.
NuExtract3 has its perks. Its smaller size and permissive license make it easier to adopt. Plus, the ability to convert documents into Markdown can fit some workflows better. But when raw extraction accuracy matters most, Lift leads the pack.
Google’s Gemini Flash 3.5 complicates the race. It slightly outperforms Lift on both field and full-document accuracy. But here’s the catch—Lift is much faster. Lift’s median latency is 9.5 seconds. Gemini Flash 3.5 takes 28.1 seconds, nearly three times longer. Speed matters when processing volumes of documents.
Cost also enters the equation. Gemini 3.5 Flash runs cost $165 per session with a 28-hour runtime, which could deter budget-conscious users. Meanwhile, operating costs for models like Fable 5 and GPT 5.5 exceed $130 for 100 problems over 10 runs. Fable 5 scores 84.5% accuracy on Google’s Android Bench, where Gemini 3.1 Pro ranks fifth behind GPT 5.4 and Claude variants.
Meanwhile, Google’s Android Bench leaderboard features a crowded field: Claude Fable 5, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.7 Code, MiniMax M3, Qwen 3.7 Plus, and Qwen 3.7 Max. Gemini 3.1 Pro holds a respectable fifth place, showing Google’s varied AI portfolio competing fiercely.
On infrastructure, Google adopted the Harbor framework for Android Bench. This underpins their model evaluation and deployment, streamlining benchmarking across devices and models.
Meanwhile, DeepSeek, a Chinese AI company, is building its own inference chip. This move aims to break dependence on American Nvidia semiconductors. DeepSeek’s token usage on OpenRouter doubled in six months after releasing its V4 model in April. Market share on OpenRouter is clearly shifting. Nearly all Chinese providers have gained ground since January while U.S. companies lost share.
This shift signals more than just model updates. It reflects the industry’s tug-of-war over AI supply chains, licensing, and performance. Datalab’s Lift shows that specialized, schema-first extraction can still dominate accuracy and speed metrics. NuExtract3 leverages flexibility and licensing to stay relevant, while Gemini Flash 3.5 offers top accuracy at a cost of speed and price.
The AI document extraction wars are far from over. But the prizes are clear: precision, speed, and cost. Lift stakes a strong claim in that arena today.
Based on
- Datalab Lift vs the Field: How a 9B Schema-First Extractor Compares with NuExtract3, LlamaExtract, Marker, and Docling — marktechpost.com
- Digital-native startups are ditching rigid databases for their agentic stacks | VentureBeat — venturebeat.com
- Google updates Android Bench with new LLMs, but Gemini still lags behind – Ars Technica — arstechnica.com
- DeepSeek takes control of its hardware | Semafor — semafor.com
