DSPy 3.5 makes LiteLLM optional and XML the default

DSPy released 3.5.0 on November 10, 2026. The maintainers say plainly that it changes prompts, defaults and APIs, and they ask users to re-evaluate optimized programs after upgrading from 3.4.1.

Prompts change even if your code does not

XMLAdapter is now the default adapter. It formats and parses predictor calls, so the prompt and the output contract change even when your Signature stays the same. API-family routing also defaults to model_type="auto" instead of an explicit chat or responses choice. If you shipped a compiled program, run your eval set again before trusting its old scores. TwoStepAdapter is deprecated.

A smaller install

LiteLLM and OpenAI's Python library are now optional. DSPy calls supported providers through the bundled lm15 engine. The release notes measure the default install at 30.9 MiB and 25 dependencies on a 3.5 development snapshot, against 206.7 MiB and 65 on 3.3.0, with fresh imports at 0.49 s instead of 1.01 s. Those are the project's numbers; we did not reproduce them.

Modules that changed status

• dspy.RLM, dspy.Flex and dspy.GEPA are no longer experimental.
• ReActV2 is now dspy.ReAct, with structured History and a typed submit tool. The old name stays deprecated until 3.6.
• Tools can return images, files and other content parts on supported provider routes.
• RLM can save and reload conversation History, but it does not restore live interpreter variables. It can also crop images with Pillow and send mixed text and image subqueries.

Safer caches, faster sandboxes

Disk caches now use restricted loading by default, and plain NumPy arrays are stored without pickle. RLM and Flex interpreters start from a precompiled standard library and cached snapshots, and InterpreterPool can prewarm them.

This entry reports the documented changes; we did not benchmark the release.

See the DSPy hub, the previous DSPy 3.4.1 note, or compare it in DSPy vs LangChain.