v2.2.0 3.4.0

Name: dspy
Author: stanfordnlp

high risk

stanfordnlp/dspySep 25, 2026 → Sep 25, 20260 days between releases

What's New

DSPy 3.4 replaces experimental LM types with native engines and introduces Jev integration via TypeSafe.

Breaking Changes

RLM interpreter API changes requiring keyword-only `interpreter_factory` option.
Experimental LM types introduced in 3.3 are replaced by native engines.
Legacy custom-LM integrations and OpenAI-style `messages=` calls are deprecated.

New Features

Jev integration with Typed Decisions (Noul, Choice, Score) for probability evidence.
Native LM engines (lm15) replacing experimental LM types.
ReAnchor optimizer for calibrating decisions against program metrics.
Async ReActV2 support.
Versioned documentation.

Upgrade Notes

Install the optional TypeSafe client with `pip install --upgrade "dspy[typesafe]==3.4.0"` and configure `dspy.experimental.TypeSafe("jev-latest")`. Update RLM interpreter calls to pass a factory via the keyword-only `interpreter_factory=` option. Note that 3.5 is the migration deadline for LM types.

The Two Releases

v2.2.0

Sep 25, 2026· 27 days ago

Engram 2.2.0 strengthens local data safety by preventing SQLite WAL usage on remote filesystems and introduces CommandCode setup alongside Windows agent hooks.

Full release notes

3.4.0

Sep 25, 2026· 27 days ago

DSPy 3.4 introduces Jev integration via TypeSafe and native LM engines, serving as a transition release for language model types while deprecating legacy integrations.

Full release notes