On-Premise AI Lab

🧪 1.3B Runs on Laptops: MiniCPM-OCR-1.0 Turns Local Invoices into JSON in Seconds: Local AI Lab

Mark Ku

October 28, 2026•14 min read

A new 1.3B parameter model, MiniCPM-OCR-1.0, recently landed on Hugging Face, enabling laptops to parse complex invoices and PDFs locally in seconds. In this issue, we will also introduce Sage-A1.5, a 35B MoE agent model that can run on a single RTX 4090. How does it help you automatically break down engineering requirements and self-correct code? We will tell you in a moment! These two open-source powerhouses, which are local-first, free, and licensed under Apache 2.0, are absolute must-haves for engineers.

Featured This Issue

1. openbmb/MiniCPM-OCR-1.0: Ultra-lightweight 1.3B Document Parsing and OCR Model, a Powerful Tool for Local Invoice and PDF Structuring

What is it: Released by OpenBMB, MiniCPM-OCR-1.0 is an ultra-lightweight vision-language model with only 1.3B parameters, specializing in document parsing and OCR tasks. It uses the developer-friendly Apache 2.0 open-source license, making it highly suitable for commercial applications. It is particularly adept at handling Traditional and Simplified Chinese and English documents, complex tables, and charts.
Use Cases:
- Scenario 1: Privacy-secure automated invoice and receipt processing. When software engineers need to develop a completely offline financial system, they can simply input images of invoices or receipts. By leveraging the model's Key Information Extraction (KIE) capabilities, they can extract the store name, transaction date, and amount directly via prompts, outputting clean JSON.
- Scenario 2: PDF document parsing pipeline for local RAG systems. When you need to build a Retrieval-Augmented Generation (RAG) system for your company and face thousands of scanned PDFs containing complex tables, mathematical formulas, and charts, you can use this model to quickly convert them into structured Markdown text, solving the formatting issues of traditional OCR.
- Scenario 3: Digitization of old or mobile-captured documents. For digital archivists or researchers dealing with skewed, blurry, or low-light mobile photos of ancient texts and historical documents, this model can accurately restore text and paragraph order, digitizing them along with chart contents.
Hardware Requirements: According to the official model card, the model has only 1.3B parameters and comes in various quantized versions such as GGUF, GPTQ, and AWQ. Developers can easily run the GGUF version directly on standard laptops or workstations. Additionally, the official release includes the DSpark speculative decoding acceleration scheme, which the documentation notes can boost end-to-end processing efficiency by about 55%, making it highly friendly for local environments with limited hardware resources.
Current Traction: It currently has a trending score of 111 on Hugging Face, with 791 downloads and 184 likes, making it a rising star to watch.
Comparison: According to OpenBMB's official tests, MiniCPM-OCR-1.0 achieves a throughput of approximately 1.2 pages per second. It leads in multiple document parsing benchmarks, such as OmniDocBench, with the official team claiming its throughput is about 1.49 times that of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro.
How to Get It: The official README indicates that you can directly load the GGUF version via llama.cpp to run a local OpenAI-compatible server using the command: llama serve -hf openbmb/MiniCPM-OCR-1.0-gguf:Q4_K_M; or deploy it using vLLM with the command: vllm serve openbmb/MiniCPM-OCR-1.0.
Link: Hugging Face Page

2. InternScience/Sage-A1.5: A 35B MoE Advanced AI Agent Model That Runs on a Single RTX 4090, Specializing in Complex Task Decomposition

What is it: Sage-A1.5 is a 35B Mixture-of-Experts (MoE) agentic model open-sourced by the Shanghai Artificial Intelligence Laboratory. Built on Qwen3.6 and licensed under Apache 2.0, it is specifically optimized for long-horizon task execution, tool calling, and agent orchestration.
Use Cases:
- Scenario 1: Autonomous Coding Agent. When software engineers develop automated code-generation tools, they need a model that can decompose complex system requirements into multiple sub-steps, automatically call external tools (such as a Code Interpreter or search engine) for testing, and even inspect error logs to self-verify and correct code when execution fails.
- Scenario 2: Local enterprise R&D and regulatory intelligence analysis. R&D personnel within enterprises need to analyze competitor intelligence or complex regulations. By leveraging the model's agentic orchestration capabilities, they can dispatch tasks to multiple parallel sub-agents to gather data and compile a comprehensive analysis report, keeping all confidential data completely local.
- Scenario 3: Exploratory research on scientific datasets. When dealing with large, complex scientific datasets, researchers need a model that can perform multi-step data analysis planning and dynamically adjust its next-step analysis strategy based on intermediate charts and data, achieving highly autonomous exploration.
Hardware Requirements: According to the official technical report and model card, although Sage-A1.5 is a 35B MoE model, only about 3B parameters are activated per token. In BF16 precision, the weight size is around 65.4 GiB, which drops to about 19.6 GiB after 4-bit quantization (Q4_K_M). Third-party evaluations indicate that under Q4_K_M quantization, the model, combined with an 8K token context length, can fit entirely within a single RTX 4090 (24 GB VRAM), bringing powerful agent capabilities to personal workstations.
Current Traction: Highly popular on Hugging Face, it currently has a trending score of 299, with 8.2k downloads and 350 likes.
Comparison: The technical report from the Shanghai Artificial Intelligence Laboratory indicates that compared to its predecessor, Agents-A1, Sage-A1.5 shows significant improvements in engineering and agentic tasks. For instance, it scored 17.6 points higher on SWE-bench Pro and 14.0 points higher on SWE-bench Multilingual. In the BrowseComp benchmark, Sage-A1.5 achieved a score of 81.7%, surpassing the previous generation's 75.5%.
How to Get It: The official README indicates that using the SGLang framework is recommended for high-performance deployment. The startup command is: python -m sglang.launch_server --model-path InternScience/Sage-A1.5 --port 8000 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder.
Link: Hugging Face Page

Summary

As we can see from this issue's selection, the open-source community is rapidly evolving toward two extremes: extreme lightweighting and powerful agentic capabilities. Whether you want to handle document structuring on your laptop with MiniCPM-OCR-1.0 or build automated agents on your workstation with Sage-A1.5, local hardware can now handle it with ease. Go ahead and pick one to download and try on your machine. See you in the next issue!