DSPy 3.4.0 released
A major release adding typed decision outputs through Jev integration, native LM engines, a local Python interpreter for code execution, and async support for ReActV2.
What changed in 3.4.0
- Jev backend now available through TypeSafe with three new decision types—Noul, Choice, and Score—that return probability evidence alongside results.
- ReAnchor optimizer tunes decision thresholds and weights against your program's metric without modifying instructions or demonstrations.
- Native LM execution preferred by default; custom HTTP providers can be registered without writing a full engine implementation.
- LocalInterpreter enables generated code to run in a persistent subprocess with state and imports preserved across calls.
- ReActV2 supports async agent calls with async predictions and tools; documentation migrated to Zensical with versioned releases.
Check before you update
- RLM call-time interpreter overrides now require keyword-only
interpreter_factory=parameter with a zero-argument factory, not a live instance. - Experimental LM types from 3.3 are replaced; legacy custom-LM integrations and OpenAI-style
messages=calls work in 3.4 with deprecation warnings but migration required by 3.5.
Summarised by AgentGid from Stanford NLP's official release notes on GitHub, as reported there; read the original before relying on it. Price, status and release dates come from our daily data.
DSPy at a glance
| What it is | A framework for programming language models with declarative modules and optimizers instead of hand-written prompts. |
|---|---|
| Latest version | 3.5.0, Nov 10, 2026. This page is about an earlier release. |
| Release pace | 1 stable release in the last 30 days; 3.4.0 came 35 days after the previous one |
| Price | Free (OSS). DSPy is open source and free to use; you pay only for the models and infrastructure it runs on. (checked Nov 4, 2026) |
| Downloads, last 7 days | 1.6M up 22.9% |
| Gid Score | 54 #12 in Agent frameworks and SDKs |