Releases·Model Release
Sarvam Trains a 3B State-Space Vision Model to Beat 100x Larger Document AI Models
In a talk, Sarvam's Krishna Prasad Srinivasan describes training a 3B state-space vision language model for Indic document AI that runs on one GPU and reportedly beats document AI models 100x larger. The model uses block-level OCR with layout and reading-order harnesses and a four-stage curriculum from 13 trillion text tokens through RL with verifiable rewards. Four months after launch it is digitizing 35 million pages for insurers, banks, and governments, addressing the near-total absence of Indian languages in machine-readable web data.
55 engagement1 sourceWed, Sep 23, 2026, 04:30 PM
AI Overview
- Sarvam's Krishna Prasad Srinivasan says under 1% of Common Crawl is Indian-language; the 3B model runs on one GPU and beats document AI models 100x larger.
- The model uses a state space backbone instead of a transformer because pages can run to 10,000 visual tokens; SSMs keep memory constant versus quadratic attention.
- Training follows four stages: 13 trillion text tokens (English, 22 Indian languages, math, code), continual pretraining on 300M image-text pairs, SFT on 100M OCR samples, then RL.
- The team bet on block-level OCR with layout and reading order harnesses in late 2025, an approach many 2026 releases have since adopted.
- RL rewards are machine-checkable unit tests on character error rate, table structure, and grammar, making verifiable-reward RL scalable for OCR.
- Four months after launch, the model is digitizing 35 million pages for insurers, banks, and governments; a public Indic benchmark spanning the 1800s to today is promised soon.
- Sarvam's moat is a data engine for languages with no labeled data, plus evals built for real-world usefulness.
Related
Releases·Model Release
MOSS-OCR: a 0.3B block-level OCR model matches MinerU 2.5 on OmniDocBench
A team trained a 0.3B-parameter OCR model called MOSS-OCR from scratch on 50M+ patent-domain samples. It achieves a 94.46 score on OmniDocBench v1.6, matching MinerU 2.5 at 25% of the parameter count. The model is block-level only, assuming layout detection and reading order are handled externally.
A team trained a 0.3B-parameter OCR model called MOSS-OCR from scratch on 50M+ patent-domain samples. It achieves a 94.46 score on OmniDocBench v1.6, matching MinerU 2.5 at 25% of the parameter count. The model is block-level only, assuming layout detection and reading order are handled externally.
17 engagement1 sourceredditMon, Jul 27, 2026, 05:29 PM
Developments·Paper
Jina AI releases Jina-OCR-v1: 3B MoE document parser hitting 2.57 pages/sec on low-budget GPUs
Jina AI introduced Jina-OCR-v1, an end-to-end document parsing model designed to run on low-budget GPUs. It reuses DeepSeek-OCR's compressed-vision encoder and 3B mixture-of-experts decoder and adds a FastMTP speculative decoding head for fast, lossless decoding. The work matters to practitioners needing cheap OCR serving without high-end hardware.
Jina AI introduced Jina-OCR-v1, an end-to-end document parsing model designed to run on low-budget GPUs. It reuses DeepSeek-OCR's compressed-vision encoder and 3B mixture-of-experts decoder and adds a FastMTP speculative decoding head for fast, lossless decoding. The work matters to practitioners needing cheap OCR serving without high-end hardware.
0 engagement1 sourcearxivWed, Sep 2, 2026, 09:49 PM
Developments·Paper
GravityOCR combines diffusion drafting and AR verification to speed up document OCR decoding
A new paper introduces GravityOCR, a parameter-shared model that combines diffusion-based parallel drafting with causal autoregressive verification to accelerate OCR inference. Because OCR outputs are strongly grounded in the input image, parallel diffusion drafting works well, and verifying drafts before commitment avoids errors from tokens predicted simultaneously. The approach delivers a 3.94x decode-only speedup while preserving OCR accuracy near the original GLM-OCR baseline.
A new paper introduces GravityOCR, a parameter-shared model that combines diffusion-based parallel drafting with causal autoregressive verification to accelerate OCR inference. Because OCR outputs are strongly grounded in the input image, parallel diffusion drafting works well, and verifying drafts before commitment avoids errors from tokens predicted simultaneously. The approach delivers a 3.94x decode-only speedup while preserving OCR accuracy near the original GLM-OCR baseline.
0 engagement1 sourcearxivTue, Sep 22, 2026, 04:11 PM
Developments·Paper
Evaluating In-Context Learning and Retrieval Strategies for Devanagari Post-OCR Correction
A new arXiv paper presents the first systematic evaluation of LLMs (3B-32B) for training-free post-OCR correction in Hindi and Marathi. The authors propose CharBM25, a character n-gram BM25 retrieval strategy for in-context examples, which outperforms random and dense retrieval on a 20,000-sentence benchmark. Gemma-3-27B achieves the best results, while models under 8B fail to reliably improve over the OCR baseline.
A new arXiv paper presents the first systematic evaluation of LLMs (3B-32B) for training-free post-OCR correction in Hindi and Marathi. The authors propose CharBM25, a character n-gram BM25 retrieval strategy for in-context examples, which outperforms random and dense retrieval on a 20,000-sentence benchmark. Gemma-3-27B achieves the best results, while models under 8B fail to reliably improve over the OCR baseline.
0 engagement1 sourcearxivFri, Sep 18, 2026, 10:26 AM
Developments·Paper
ArXiv paper identifies four design principles for pixel-space text representation learning
A new arXiv paper studies why pixel-text encoders fail at reading, retrieving, and compressing language directly in pixel space. Through controlled ablations it identifies four critical design components for robust visual text representation. The findings target problems like fixed-resolution pretraining, visual shortcut learning, weak grounding, and multilingual visual text understanding.
A new arXiv paper studies why pixel-text encoders fail at reading, retrieving, and compressing language directly in pixel space. Through controlled ablations it identifies four critical design components for robust visual text representation. The findings target problems like fixed-resolution pretraining, visual shortcut learning, weak grounding, and multilingual visual text understanding.
0 engagement1 sourcearxivTue, Sep 1, 2026, 12:20 PM