Reported News
DSPy 3.5 makes LiteLLM optional, graduates RLM, Flex and GEPA, and switches the default adapter to XML
DSPy 3.5.0, released on GitHub on November 10, 2026 (UTC), "brings a lighter LM layer, graduated core modules, and new History and image capabilities for RLM," per its release notes. The notes report the measured default install falling from 206.7 MiB in 3.3.0 to 30.9 MiB, and warn that the release changes prompts, defaults and APIs.
- Changes reported
- 6
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 4
- things we could not verify
- Sources
- 1
- 1 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | LiteLLM and OpenAI's Python library are now optional: DSPy calls supported providers through a lightweight engine called lm15. The notes measure the default install at 30.9 MiB and 25 dependencies, against 206.7 MiB and 65 in 3.3.0, and fresh-process import at 0.49 s versus 1.01 s. | Reported |
| 02 | dspy.RLM, dspy.Flex and dspy.GEPA are no longer experimental, and ReActV2 is now simply dspy.ReAct, with structured History and a typed submit tool; the old name is deprecated until 3.6. | Reported |
| 03 | XMLAdapter is now the default adapter, which the notes say changes prompts and the output contract even when a Signature stays the same. model_type="auto" is also the new default for choosing between chat and responses APIs. | Reported |
| 04 | RLM can save and reload conversation History (without restoring live interpreter variables), crop images with Pillow and send mixed text and image subqueries from its Python REPL; tools can return images, files and other content parts on supported provider routes. | Reported |
| 05 | Prewarmed interpreters: the notes report cached Deno startup plus first execution dropping from 1.59 s in 3.4.1 to 0.45 s, and a Flex forward with a ready size-one InterpreterPool at 0.10 s. | Reported |
| 06 | Disk caches are secure by default: restricted loading is now the default and plain NumPy arrays are stored without pickle. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: GitHub release: stanfordnlp/dspy 3.5.0. We are the fast, independent index; the vendor’s own page is the source of record for this release.
Who should act before upgrading
Anyone running optimized DSPy programs in production should treat this as a breaking upgrade. The default adapter switch to XML changes the prompt text and the output format the model is asked to follow, so prompts compiled by an optimizer under 3.4 may score differently. The release notes themselves tell users to read the upgrade guide and re-evaluate optimized programs.
Teams that depended on LiteLLM features through DSPy should check whether their provider is covered by the new lm15 path; the notes say LiteLLM is now optional rather than removed, and keep a compatibility route.
What the release means for builders
A nearly sevenfold smaller default install and a halved import time, as the notes measure them, matter most for serverless functions and containers where cold starts are billed. Graduating RLM, Flex and GEPA out of experimental status signals the maintainers consider those APIs stable enough to build on.
In the source’s words
“DSPy 3.5 brings a lighter LM layer, graduated core modules, and new History and image capabilities for RLM.”
“This release changes prompts, defaults and APIs.”
“changing prompts and the output contract even when your Signature stays the same”
What we have not verified
- We did not run this release ourselves.
- The install size, dependency count, import time and interpreter startup figures are the maintainers' own measurements, partly on a 3.5 development snapshot; we did not reproduce them.
- We did not check which providers the lm15 engine supports or where LiteLLM is still needed.
- We did not measure how much the XMLAdapter default changes the scores of previously optimized programs.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- GitHub release: stanfordnlp/dspy 3.5.0primary source · fetched 2026-11-11
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/dspy-3-5-lighter-lm-layer-xmladapter-default.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.