• AI Engineering
  • Posts
  • IBM released Granite-Docling-258M Model for Document AI

IBM released Granite-Docling-258M Model for Document AI

.. PLUS: Memory for AI Agents in 6 lines of code

Author

Sumanth P
September 29, 2025

In today’s newsletter:

  • Cognee - Dynamic memory for AI Agents

  • IBM Granite-Docling-258M: Tiny Multimodal Model for Document AI

  • AgentScope – Alibaba’s Framework for Multi-Agent AI Apps

Reading time: 3 minutes.

Cognee let's you build memory for Agents and replace RAG using scalable, modular ECL (Extract, Cognify, Load) pipelines.

Sending large volumes of data to AI agents often leads to bloat and hallucinations.

Cognee connects data points and establishes ground truths to improve the accuracy of your AI agents and LLMs.

Key Features:

  • Interconnect and retrieve your past conversations, documents, images and audio transcriptions

  • Replaces RAG systems and reduces developer effort, and cost.

  • Load data to graph and vector databases using only Pydantic

  • Manipulate your data while ingesting from 30+ data sources

  • Local UI with interactive notebooks for easy data loading, graph visualization, and querying

It also supports continuous improvement through a feedback mechanism that captures the relevance of search results from real user interactions.

Over time, this feedback directly updates the knowledge graph, helping your agents adapt and provide increasingly accurate responses.

It's 100% Open Source

IBM released Granite-Docling-258M, a compact vision-language model built for end-to-end document conversion.

It turns complex documents into structured data while preserving layouts, tables, equations, and more all at a fraction of the compute cost.

Key Features:

  • Ultra-Compact -258M parameters with performance rivaling larger models.

  • DocTags Format - Structured markup with coordinates to capture tables, equations, code, and captions.

  • Math Recognition - Parses inline and block formulas with high accuracy.

  • Flexible Modes - Full-page or bbox-guided region conversion.

  • Stability - Reduced errors and token loops.

  • Multilingual - Early support for Arabic, Chinese, Japanese, and Latin scripts.

It’s 100% Open Source.

Alibaba released AgentScope, a Python framework for building modular, multi-agent LLM applications.

Built on the concept of Agent-Oriented Programming (AOP), which treats agents as first-class components, making it easier to design, control, and scale AI systems.

The framework focuses on giving developers full transparency and runtime control, while still being flexible and model-agnostic. Instead of black-box orchestration, every step of reasoning, tool usage, and workflow remains visible and interruptible.

Key Features:

  • Transparency - Full visibility over prompts, reasoning chains, and agent interactions.

  • Runtime Control - Native support for interruptions, overrides, and custom error handling.

  • Agentic Core - Built-in handling for tools, memory, RAG, and multi-agent collaboration.

  • Model Agnostic - Works with any LLM provider without lock-in.

  • Composable - Modular, LEGO-style components let developers mix and match agents.

It’s 100% Open Source.

That’s a Wrap

That’s all for today. Thank you for reading today’s edition. See you in the next issue with more AI Engineering insights.

PS: We curate this AI Engineering content for free, and your support means everything. If you find value in what you read, consider sharing it with a friend or two.

Your feedback is valuable: If there’s a topic you’re stuck on or curious about, reply to this email. We’re building this for you, and your feedback helps shape what we send.

WORK WITH US

Looking to promote your company, product, or service to 150K+ AI developers? Get in touch today by replying to this email.