TeXada Releases Local Math Agent Built on MiniCPM

TeXada arrived as a community project that turns natural language and images into editable LaTeX using only MiniCPM5-1B and MiniCPM-V 4.6 running on local hardware.

The release focuses on three tasks. Users type a problem in plain English and receive valid LaTeX. They photograph handwritten equations or diagrams and obtain clean editable code. The agent also completes partial expressions and corrects common syntax errors before a document reaches a compiler.

All processing stays on the device. No prompts leave the machine, which removes the usual privacy risk when mathematical work contains proprietary data.

Students benefit during exams or late-night study sessions where internet access is restricted. Researchers keep sensitive derivations inside institutional networks. Developers embed the agent in documentation pipelines that must remain offline for compliance reasons.

The first public repository appeared on GitHub with model weights also hosted on Hugging Face. Early testers reported that handwritten integral signs and matrix notation survive the OCR stage with fewer than five percent symbol errors on standard phone cameras.

MiniCPM5-1B handles the language to LaTeX mapping after fine-tuning on synthetic math corpora. MiniCPM-V 4.6 manages the vision side, converting pixel input into token sequences that the language model then refines. Both models fit on consumer laptops with 16 GB of unified memory.

The project highlights a narrow but practical use case. Most frontier model APIs still require an internet connection and send user equations to remote servers. TeXada removes that hop entirely for one domain.

Local execution imposes limits. Context windows remain shorter than those available in larger hosted models, so very long proofs must be broken into smaller sections. The agent does not yet attempt symbolic solving beyond pattern completion; it focuses on representation rather than derivation.

Three trade-offs surface quickly. Speed stays acceptable on recent Apple silicon and mid-range NVIDIA cards but slows on older CPUs. Accuracy on highly stylized handwriting varies with training data quality. The open weights invite further fine-tuning yet also expose the models to misuse if someone adds malicious instruction tuning.

A small set of contributors continues to publish example notebooks that show end-to-end workflows from phone photo to compiled PDF. One notebook demonstrates live correction inside a VS Code extension that calls the local inference server.

Future updates listed in the repository roadmap include support for multi-line aligned environments and better handling of chemical structure notation. No timeline has been announced for those additions.

The release occurs against a backdrop where several larger labs have moved more math capabilities behind paid APIs. A fully local alternative therefore fills an immediate gap for users who cannot accept recurring cloud costs or data transfer policies.

Early GitHub issues already list requests for better support of commutative diagrams and faster batch processing of assignment sets. The maintainers have marked several of those requests as open for community pull requests.

TeXada demonstrates that targeted fine-tuning of small multimodal models can deliver usable results in a constrained domain without external dependencies. The approach trades breadth for guaranteed privacy and zero marginal cost after the initial download.

Readers who need occasional LaTeX conversion without leaving their local environment now have a concrete option that runs entirely on their own hardware. The repository remains the primary source for the current weights and extension code.