# Tesseract-JAX v0.4.0 released

**URL:** <https://si-tesseract.discourse.group/t/tesseract-jax-v0-4-0-released/142>\
**Category:** Release Notes\
**Created:** [June 25, 2026, 6:15pm UTC](https://si-tesseract.discourse.group/t/tesseract-jax-v0-4-0-released/142 "2026-06-25T18:15:19Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![dionhaefner](https://yyz1.discourse-cdn.com/flex009/user_avatar/si-tesseract.discourse.group/dionhaefner/32/13_2.png) [@dionhaefner](https://si-tesseract.discourse.group/u/dionhaefner)\
**Post date:** [June 25, 2026, 6:15pm UTC](https://si-tesseract.discourse.group/t/tesseract-jax-v0-4-0-released/142/1 "2026-06-25T18:15:20Z")

</div>

# Release v0.4.0

## Action required

### Python 3.10 is no longer supported; JAX ≥ 0.7.0 required (#207)

The minimum supported Python version is now **3.11** , and the minimum JAX version is now **0.7.0**. Python 3.10 is nearly end-of-life and JAX dropped support for it in 0.7.0.

To upgrade, make sure your environment uses Python 3.11+ and a recent JAX, then update Tesseract-JAX as usual:

```bash
pip install --upgrade "tesseract-jax"

```

CI now also tests against Python 3.14.

## Highlights

### Much faster Jacobians via the `jacobian` endpoint (#191)

`jax.jacfwd`, `jax.jacrev`, and `jax.jacobian` over an `apply_tesseract` call (as well as `jax.vmap` over tangents/cotangents with unbatched primals) now dispatch to the Tesseract `jacobian` endpoint and contract the materialized Jacobian in a single pass, instead of calling the `jvp`/`vjp` endpoint once per batch element.

No code changes are required — existing transformations automatically take the fast path when the Tesseract exposes a `jacobian` endpoint. In benchmarks this is **14x faster for a 100×10 Jacobian and ~140x faster for a 100×100 Jacobian**.

This unlocks differentiable-systems applications such as least-squares solves and root finding with JAX-based optimizers, as well as sensitivity analysis:

```python
import jax
from tesseract_jax import apply_tesseract

# `jacfwd` / `jacrev` now use the jacobian endpoint under the hood
jac = jax.jacfwd(lambda x: apply_tesseract(tess, {"a": x})["result"])(x0)

```

## Other changes

- Fixed random deadlocks caused by allocating JAX memory from within Python callbacks; callbacks now return NumPy arrays only (#177).
- Bumped the lowest _runtime-tested_ JAX version to 0.6.2 internally before the 0.7.0 requirement landed (#201).
- Docs: pulled in the new navbar from tesseract-core (#183) and corrected the comment on the VJP non-differentiable input placeholder (it is NaN, not zero) (#190).

## What’s Changed

### Features

- Use jacobian endpoint (#191)
- _(deps)_ [**breaking**] Drop support by Python 3.10, test Python 3.14 in CI, bump jax lb to 0.7.0 (#207)

### Bug Fixes

- Prevent random deadlocks by not allocating JAX memory from within callbacks (#177)

### Refactor

- Bump lowest supported JAX version to 0.6.2 (#201)

### Documentation

- Pull in new navbar from tesseract-core (#183)
- Correct comment on VJP non-diff input placeholder (NaN, not zero) (#190)

**Full diff** : [Comparing v0.3.0...v0.4.0 · pasteurlabs/tesseract-jax · GitHub](https://github.com/pasteurlabs/tesseract-jax/compare/v0.3.0...v0.4.0)