Surpassing 10,000 downloads right at launch, Xiaomi-OCR-0, a 0.8B ultra-lightweight model, delivers stunning performance in complex document parsing! In this issue, we have selected two open-source gems with extremely low entry barriers. Besides this invoice structured parsing expert, we also feature Edge0/Audio8-TTS-Realistic-0.6b, which can clone multilingual voices using just 1.3 GB of VRAM. What on-premise pain points can they solve? We'll tell you in a moment!

---

Featured Projects

1. xiaomi-research/Xiaomi-OCR-0: A 0.8B Ultra-Lightweight Document Understanding and Structured Parsing Model

• What it is: This is a 0.8B ultra-lightweight document understanding model released by the Xiaomi Research team (xiaomi-research). Built on the Qwen3.5-0.8B-Base foundation model and licensed under the developer-friendly Apache-2.0 open-source license, it is highly suitable for high-precision local image text extraction and structured parsing.
• Use Cases:
  1. Automated Bookkeeping and Expense Reporting: Software engineers can automatically extract specific fields (such as amounts, dates, and tax IDs) from large volumes of PDF bills or invoice images and convert them into structured JSON format. This allows direct import into accounting systems without worrying about uploading confidential financial data to the cloud.
  2. Privacy-Secure Visual Question Answering (VQA): Developers looking to build a highly secure document Q&A system locally can allow users to upload contracts or internal report images directly and query data details using natural language, ensuring sensitive business contracts never leave the local environment.
  3. Digitizing Historical Archives and Complex Layouts: System integrators can parse historical document images with complex layouts, handwriting, or tables, converting them into structured Markdown or plain text to facilitate the creation of full-text search systems.
• Hardware Requirements: According to the official model card, the team has not specified a minimum VRAM requirement, but they note its 0.8B parameter scale with a download size of about 1.8 GB. The official documentation also mentions that the actual hardware resources required depend on the vision encoder, image resolution, and context length. Thanks to its extremely small parameter size, it is highly accessible for developers running it locally on low-spec machines. The currently available format is safetensors.
• Current Traction: It currently has a trending score of 97 on Hugging Face, with approximately 16.0k downloads and 153 likes.
• Comparison: According to the official technical report, the model scores 95.24 on Real5-OmniDocBench and averages 83.2 across five OCR-oriented VQA benchmarks. This performance outperforms Qwen3.5-0.8B (72.9) and Qwen3.5-2B (80.9), making it highly competitive among sub-1B models.
• How to Get Started: The official README indicates that it can be deployed via vLLM using the following command:

  vllm serve xiaomi-research/Xiaomi-OCR-0 --host 127.0.0.1 --port 8000 --max-model-len 16384
  

  Alternatively, you can load it directly using Transformers.
• Link: xiaomi-research/Xiaomi-OCR-0

---

2. Edge0/Audio8-TTS-Realistic-0.6b: A Multilingual Speech Synthesis and Voice Cloning Model Running on Just 1.3 GB of VRAM

• What it is: This is a 0.6B ultra-lightweight speech synthesis and voice cloning model released by Edge0, also under the Apache 2.0 license. It delivers highly realistic multilingual voice generation and voice cloning with minimal hardware resources.
• Use Cases:
  1. Indie Game NPC Voice Acting: Indie game developers on a tight budget who want to add multilingual voice acting for in-game NPCs can generate natural, fluent speech locally by providing just a 3-second reference audio clip and the script, saving significant voice-acting costs.
  2. Automated Podcast and Video Voiceovers: Podcast creators or video editors who need to quickly convert text scripts into voiceovers can generate speech with natural rhythm, pacing, and breathing pauses rather than a robotic machine voice, providing a more immersive experience for listeners.
  3. Voice Assistants on Low-Spec Edge Devices: Software engineers looking to build a local voice assistant supporting 23 languages can achieve extremely low inference latency and memory footprint, making it easy to integrate into low-spec edge devices.
• Hardware Requirements: According to the official README, it requires only about 1.3 GB of VRAM in FP16 inference mode, and INT4 quantization can reduce the memory footprint even further. Additionally, the paper notes that running the INT4 ONNX format on an Apple M2 CPU requires a peak memory of only 1.1 to 1.2 GiB, making it extremely friendly for developers without high-end GPUs or those working solely on a Mac. The currently available format is safetensors.
• Current Traction: It currently has a trending score of 140 on Hugging Face, with 2.8k downloads and 142 likes.
• Comparison: According to official and community benchmarks, the model achieves an average error rate of just 4.86% on CV3 (covering 9 languages), outperforming CosyVoice3-1.5B (6.89%) and Qwen3-TTS-1.7B (7.73%). It delivers performance comparable to or even better than much larger models, despite its tiny 0.6B size.
• How to Get Started: The official README states that in a Python 3.10+ environment with a compatible GPU, you can load it directly using Transformers with the following code:

  from transformers import AutoProcessor, AutoModel
  
  processor = AutoProcessor.from_pretrained("Edge0/Audio8-TTS-Realistic-0.6b", trust_remote_code=True)
  model = AutoModel.from_pretrained("Edge0/Audio8-TTS-Realistic-0.6b", trust_remote_code=True)
  

• Link: Edge0/Audio8-TTS-Realistic-0.6b

---

Summary

As we can see from today's selection, "small-parameter, high-precision" edge models are rising rapidly. Whether in document parsing or speech synthesis, models under 1B parameters are delivering stunning performance that surpasses older, larger models. If you happen to have projects involving invoice bookkeeping or contract parsing, we highly recommend trying Xiaomi-OCR-0 first. You can build a secure, leak-proof local workflow with minimal resources. See you in the next issue!