AI library

Source · release

DSPy 3.4.0 release notes

Author
Stanford NLP (DSPy maintainers)
Published
25 September 2026
Added
12 October 2026
Original
https://github.com/stanfordnlp/dspy/releases/tag/3.4.0

Summary

DSPy 3.4.0 (2026-09-25, date from PyPI) is the "LM transition" release of the framework for programming LLM pipelines and optimising their prompts automatically. It moves model calls onto a native engine layer (dspy.lm15), with LiteLLM as a fallback, and deprecates the legacy interfaces, which will be removed in 3.5. The 3.x line started with 3.0.0 on 2025-08-12.

Key points

  • Native LM engines: engine="auto" prefers native execution and falls back to LiteLLM; custom HTTP providers and engines can be registered.
  • Typed decisions (experimental): through a TypeSafe integration, new output types Noul (boolean), Choice, and Score return values with probabilities and confidence. ReAnchor tunes their thresholds against a metric without changing prompts.
  • Agents and code: async ReActV2 (acall); a LocalInterpreter subprocess for RLM and Flex, explicitly not a security sandbox.
  • GEPA optimiser: a code_proposer= hook for custom program-structure proposals (Flex, from 3.3, lets GEPA change program structure, not just prompts).
  • Breaking changes: the 3.3 experimental typed request/response types are removed; Pydantic 2.11 is the minimum; native n answers now send sequential requests and can bill input tokens several times.
  • Deprecated, removed in 3.5: OpenAI-style lm(messages=[...]) calls and custom BaseLM.forward() integrations. No date for 3.5 was given.

Why it matters

DSPy remains the main framework built around automatic prompt and program optimisation. This release cuts its dependency on LiteLLM and shows optimisation extending from prompts to program structure and decision thresholds.