Mistral Large 4 Weights Are Out + Surya 2.5 Speeds Up PDF and Document OCR

AI and SaaS Roundup · 2026-10-31

Mistral’s big coding model is now downloadable, and there’s a much smaller release for turning messy documents into usable text and tables. Two very different ways to build more into your own tools. 👀

Mistral Large 4 is now available to download

Mistral released the improved Large 4 weights on Thursday, following the public preview earlier this month. It takes text and images, supports tool calls and structured JSON, and gives you low, medium or high reasoning settings. Possible uses include coding agents and assistants that work through charts or documents.

There are official FP8 and NVFP4 checkpoints, plus GGUF versions from the community. This is a roughly trillion-parameter model, so downloading it doesn’t mean it’ll comfortably run on your laptop. The published serving examples use multiple GPUs; the Mistral API is another way to access it.

The licence detail matters if you’re building a business around it: the Mistral Ecosystem Open-Weight License allows commercial and non-commercial use, but operating a model-as-a-service business requires a separate commercial licence. Open weights, with that restriction.

Source: Mistral

Surya 2.5 makes self-hosted document OCR faster

Datalab released Surya 2.5 on Thursday: a 689M-parameter model for reading documents into text, tables and layout information. Useful building blocks for a searchable document library, a PDF import feature or a document-processing workflow.

Datalab reports 1.7 times the page throughput of Surya 2 in its H100 test. It also published Apple Silicon results with llama.cpp, so there’s a route for running it on a Mac. Those are the team’s test results, not a speed guarantee for every document.

The surya-ocr package uses the new model by default from version 0.23.0. One limitation: tables come back as HTML, without row and column coordinate boxes.

The code is Apache 2.0, but the weights use a modified OpenRAIL-M licence. Datalab describes free use for research, personal projects and startups below its $5M funding/revenue threshold; broader commercial use needs separate licensing.

Source: Datalab