The Tencent n 1 CI2 model, released by the Prisma LM team, is available in three versions: 2, 4, and 9 billion parameters. The versions are designed to run on consumer systems. The article examines how the model processes document images, removes personal data, and passes the cleaned data on to the downstream processing pipeline.

Key Model Characteristics

The model accepts text queries and images. The system uses flexible visual data processors to analyze all visual content.

Loading the Model

Model files are published on the official Hugging Face page. The choice of version depends on the hardware configuration:

• for a graphics card with 2 GB of memory, the GGML files of the 2-billion-parameter version are suitable;
• systems with greater resources can use larger versions.

In testing, 2-bit quantization was used. Additionally, the vision module file was downloaded. The download was performed in the Google Antigravity development environment using hf download commands into the local models folder.

Registering the Model in Ollama

GGML files are registered with the Ollama service using a simple custom model file. The model is given the name ternary bonsai two. To interact with it, the ollama.chat command is used with the specified model name. To install Ollama, download the free software from ollama.com/download.

Operating Mechanism

Architecture and Computation

• Three numerical states. The model stores its full knowledge base in three values. Compressed data is not decompressed: computations are performed directly in GPU memory.
• Addition instead of multiplication. Resource-intensive multiplication logic is excluded from the core. Computations rely on simple addition, which reduces the load on computing resources.
• Hybrid attention architecture. Layers of standard and linear attention enable high-speed processing of long inputs.
• Mathematical smoothing. A dedicated smoothing process handles extreme values and preserves output quality during compression.

Image Processing Module

The image processing module remains inactive until it is needed. Even when active, it occupies a small amount of memory. For text queries, the module is bypassed. The model supports dense documents and recognizes small text, numbers, and handwritten data.

Document Processing Pipeline

The script main.py performs the following actions:

1. Reads all images from the inputs folder.
2. Passes the images and the recognized text to Ollama along with the model name.
3. Saves the JSON response to the outputs folder.

Request Requirements

The model is given the following task:

• find personal data: names, contact numbers, email addresses;
• identify payment amounts and sensitive financial data;
• produce a response in the specified format, including:
  • a document summary;
  • a list of the redacted text indicating the hidden information;
  • a classification of all data by entity type;
  • an audit log.

Test Data

The test set included five documents:

1. A hotel receipt.
2. A café receipt.
3. A company business card.
4. A complex invoice with handwritten text and signatures.
5. A medical prescription in a non-English language.

Результаты тестирования

Аппаратные показатели

• Использовано 0,8 ГБ памяти видеокарты из 4 ГБ доступных.
• Обработка пяти документов заняла 30 секунд, примерно 6 секунд на документ.

Документ 1: чек отеля

• Отображены название отеля и адрес.
• Имя владельца скрыто.
• Выведены стоимость позиций, итоговая сумма и налог 7,6%. Значения 54,5 и 7,6% указаны корректно.
• Скрыты имя сотрудника, телефон, факс и адрес электронной почты.
• Предоставлен полный журнал скрытых данных.

Документ 2: чек кафе

• Скрыты имена и фискальный идентификатор.
• Указаны позиции и их стоимость.
• Промежуточная сумма отсутствует в чеке. Модель рассчитала ее самостоятельно: 1,8 + 1,8 = 3,6, итог 3,6.
• Идентификатор счета скрыт.
• Предоставлен журнал аудита.

Документ 3: визитная карточка

• Название: Sleepy Hollow Farm, специализация: морские минералы и сельское хозяйство.
• Скрыты имена обоих владельцев и их мобильные номера.

Документ 4: исторический счет

• Модель определила документ как счет ювелирной компании Imperial 1900 года, выставленный императору Николаю II.
• Извлечены адрес головного офиса и сведения о частных казначейских счетах.
• Указаны магазины компании: Санкт-Петербург, Москва и Одесса.
• Скрыты имя клиента и номер счета.
• Дата: 10 мая 1902 года.
• Итоговая сумма: около 14 800 рублей.
• Основные позиции извлечены корректно.
• Модель пропустила часть подписей и имен.

Документ 5: медицинский рецепт

• Модель определила документ как итальянский документ туберкулезного диспансера, описывающий протоколы госпитализации детей.
• Название извлечено корректно.
• Имя пациента и идентификатор протокола скрыты.
• Дата отображена как 10 10 61. В итоговом извлечении эта дата отсутствует.
• Текст печати извлечен, подпись скрыта.
• Модель пропустила часть числовых данных.

Общая оценка

Модель в целом справилась с задачей. На сложных документах фиксируются пропуски отдельных данных. Обработка каждого документа заняла 5–10 секунд при полностью локальном запуске.

Benchmark Results

According to the material, the model showed the following results:

• higher results in standard tests on elementary-school-level mathematical problems;
• a 66% increase compared to the base model Q1;
• retention of the ability to call external functions and system tools;
• a significant increase in generation speed on modern laptop processors;
• lower battery charge consumption;
• retention of most of its capabilities for complex mathematical reasoning;
• superiority over the base model in basic mathematics, instruction following, and programming tests;
• retention of about 98% of the base model's intellectual capabilities in tool-calling and multimodal analysis tasks.

These results are presented in the material without independent verification.

Use Cases

• Data anonymization. Developers use the model to remove confidential information from documents. The model runs fully offline, so customer data and details never leave the system.
• Interactive voice response systems (IVR). The model can be used in such systems.
• Operation on low-power hardware. The model runs on efficient processors without overheating and requires low power consumption.
• Mobile devices. The model can read screen content and interact with applications without an internet connection.
• Edge devices with limited memory. Quantized versions under 1 GB are suitable for standard laptops, phones, and resource-constrained devices.
• Vision tasks. For image processing, a separate vision module is connected to the model.

Advantages and Limitations

Advantages

• Runs on low-power hardware, including older laptops, without crashes or freezes.
• Processes documents without cloud servers, with full control over personal data.
• Generation speed of up to 310 tokens per second on consumer laptops.

Limitations

• The model may produce lengthy internal reasoning. For quick answers, it is recommended to switch the model to fast-response mode.
• The generation temperature should be set to deterministic values. This reduces the risk of repetitions and logical loops.
• For maximum performance, a separate backend based on the llama.cpp runtime is used.
• On complex documents, data may be omitted, including numbers, signatures, and some names.