This issue features indra-ocr, an ultra-lightweight OCR model with only 1.25M parameters that processes Traditional Chinese OCR at high speeds on CPU servers without needing a GPU, and Kromoss, an image generation model that can quickly generate clear Chinese text in ComfyUI in just 8 steps. Are you a backend engineer looking to protect customer privacy without paying for expensive cloud APIs? We will show you how to build a high-throughput document processing pipeline locally using the most resource-efficient architecture.
Featured This Issue
1. datalab-to/indra-ocr: Ultra-lightweight, CPU-first, blazing-fast OCR model supporting Traditional Chinese and LaTeX formula conversion
- What it is: This is an ultra-lightweight, CPU-first OCR model with only 1.25M parameters, licensed under OpenRAIL. It is specifically optimized for Traditional Chinese and English, allowing you to quickly convert scanned documents, receipts, or academic PDFs into Markdown or structured JSON formats.
- Use Cases:
- Backend engineers building fully local document processing pipelines: When you need to automatically recognize and convert scanned invoices and receipts containing Traditional Chinese and English into structured JSON format to fully protect customer privacy, this model eliminates any compliance worries about sending data to the cloud.
- Full-stack developers building full-text search for historical PDFs: If you need to build a search index for a massive archive of internal scanned documents but have a tight budget and want to avoid expensive GPU resources, this model can be deployed directly on existing CPU servers for high-throughput text extraction.
- Researchers converting academic papers: If you have scanned papers containing complex mathematical formulas that need to be converted into editable Markdown format, you can use the
--mathparameter to let the model automatically convert formulas into precise LaTeX syntax.
- Can it run: The official README labels it as CPU-first. The model size is extremely compact, with only about 1.25M parameters, and the ONNX pipeline is only around 18 MB. According to official data, on a 4-core vCPU x86 machine, the median processing time per page is just 212 milliseconds, or 357 milliseconds with math formula mode enabled, requiring absolutely no GPU involvement. Available formats include safetensors and ONNX.
- Popularity: It currently has a trending score of 255 on Hugging Face, with 3.1k downloads and 285 likes.
- Comparison: According to the official model card and third-party benchmarks, its processing speed can reach 44 pages per second on a 32-core AMD EPYC server, which is up to 15 times faster than Tesseract 5 and 10 times faster than PaddleOCR (PP-OCR). However, third-party evaluations also note that its error rate may increase when dealing with extremely old or blurry scanned books.
- How to get it: According to the official README, developers can install it directly via
pip install indra-ocr. Runindra-ocr paper.pdfin the terminal to output Markdown to stdout, or runindra-ocr scan.png --format jsonto quickly get a JSON format containing lines, bounding boxes, and layout blocks. - Links: Hugging Face Page Link
2. lodestones/Kromoss: Bilingual ComfyUI image generation model for fast, 8-step local generation of clear Chinese text
- What it is: This is a community full-finetuned version based on MOSS-Image-Base with around 5B parameters, licensed under Apache-2.0. Its standout feature is support for bilingual Chinese and English prompts, and it comes with an 8-step DMD/DMD2 distilled LoRA, significantly reducing the time required to generate high-quality images with Chinese text locally.
- Use Cases:
- Frontend engineers quickly generating UI mockups and prototypes: When you are designing web prototypes and need to quickly generate UI mockups with specific Chinese/English typography and structured layouts (such as flowcharts or kanban boards on a whiteboard), you can use JSON format to precisely control the placement of objects and text in the image.
- Indie hackers creating localized product promotional illustrations: If you need to create promotional images with Traditional Chinese handwritten fonts for your local projects, such as a detailed illustration of a fridge magnet that says "Remember to buy milk and eggs!", this model can seamlessly integrate the Chinese text directly into the image, eliminating the need to manually add text using Photoshop.
- Can it run: The official README lists the model size at around 5B parameters in safetensors format. According to a third-party GPU evaluation site, running single-batch inference at FP16 precision requires an estimated peak VRAM of about 11 GB, which can be easily handled by consumer-grade graphics cards like the RTX 4090 24GB. Additionally, you can quantize it to INT8 (requiring about 5.5 GB VRAM) or INT4 (requiring about 2.8 GB VRAM) to further lower the local VRAM threshold.
- Popularity: It currently has a trending score of 162 on Hugging Face, with 5.4k downloads and 163 likes.
- Comparison: The official model card notes that this model is a full-finetuned version of MOSS-Image-Base. Compared to the official MOSS-Image-SFT, Kromoss is a seamless drop-in replacement. Furthermore, it comes with an 8-step DMD/DMD2 distilled LoRA, shortening the generation process from the original 50 steps to just 8 steps, greatly improving local inference efficiency.
- How to get it: According to the official README, when using it in ComfyUI, you need to place
kromoss_bf16.safetensorsinto themodels/diffusion_models/folder, and pair it with theqwen3.5_9b_bf16.safetensorstext encoder andflux2-vae_small.safetensorsVAE. To enable fast 8-step generation, load8-steps-dmd-lora.safetensorsand set the KSampler to euler/simple, 8 steps, and CFG 1. - Links: Hugging Face Page Link
Conclusion
As seen from this issue's selection, local AI models are rapidly evolving in two directions: extreme lightweighting and high localization. Whether it is indra-ocr, which runs at high speeds on CPU servers without even needing a GPU, or Kromoss, which perfectly generates Traditional Chinese handwritten text, both demonstrate the open-source community's incredible creativity in building practical tools. If you happen to be facing challenges with document recognition or UI mockup generation, why not download these models to your machine this weekend and give them a try? See you in the next issue!
Comments