Model guideSources checked October 14, 2026
Matilda Jev
Matilda Jev is Maincode’s decision model, separate from TypeSafe AI’s Jev. It offers downloadable weights and supports text, JSON, and optional image inputs. Running it requires substantial hardware: the published v1 bf16 weights alone need about 49 GiB before runtime overhead. Before trying Matilda Jev, check the exact release, serving requirements, and what its reported benchmark results actually measure.
Matilda Jev and TypeSafe Jev: keep the names separate
| Name | Publisher | What to look for |
|---|---|---|
| Jev | TypeSafe AI | TypeSafe’s documentation and hosted model identifiers |
| Matilda Jev | Maincode | The specific Maincode repository, release, and serving instructions |
TypeSafe identifies Jev as its own System One model. Maincode publishes Matilda Jev separately. Do not assume that a tutorial, benchmark, or configuration for one applies unchanged to the other. TypeSafe · Maincode
The repository Maincode/matilda-jev-v1 describes a model that evaluates supplied options for Choice, Score, and yes-or-no questions. The publisher and repository matter when you choose which model to download or evaluate.
What does running Matilda Jev involve?
For matilda-jev-v1, the published card lists 26.1 billion parameters and approximately 49 GiB for bf16 weights before runtime overhead. This is a substantial deployment requirement.
Maincode provides a bundled runtime. Its documentation notes that loading the backbone through AutoModel alone returns hidden states; the decision runtime also needs the readout and associated processing. Matilda Jev deployment instructions
Before renting hardware, decide what you want to learn. If the aim is simply to understand typed decisions, setting up a large model may be more work than the experiment needs. If the aim is to evaluate a self-hosted candidate, the serving setup becomes part of the assessment.
Does Matilda Jev beat Jev?
The v1 card reports a public Decision Index 0.3 score of 60.16 and marks the result as awaiting official submission. That is a version-specific, publisher-reported result, as checked on October 14, 2026. Published Matilda Jev results
Do not turn that number into a claim that Matilda Jev is better for every application.
Before comparing two scores, check that they refer to the same benchmark version, evaluation setup, and model release. If an article discusses a newer Matilda version, do not attach its result to the v1 download.
A useful comparison should also explain what happens when the model is wrong. A small gain in an overall score may matter less than avoiding a particular mistake in your workflow.
How should you test Matilda Jev?
Choose one task before downloading the model. For example, use a set of messages your team has already assigned to review categories.
Keep a record of the exact release and serving configuration. Run the same examples through your existing approach, then compare:
- Cases both systems answer correctly.
- Cases Matilda Jev improves.
- Cases Matilda Jev makes worse.
- Cases where the expected answer is itself unclear.
For image-based tasks, inspect whether the answer depends on details that are legible in the image. Include incomplete and poor-quality inputs instead of evaluating only clean examples.
These are recommended evaluation steps, not results from a UseJevAI benchmark.
Is Matilda Jev right for your project?
Matilda Jev is worth investigating if you are evaluating self-hosted decision models and can support the required infrastructure.
If you want a quick answer to “what is this model?”, start with its publisher, exact release, and deployment requirements. If you want to adopt it, follow that with a task-specific evaluation. The shared word “Jev” is a reason to check the details, not a reason to assume the models behave alike.