# DSPy 3.4.1: what's new, verdict & setup Canonical: https://theagenticweekly.com/tools/dspy · Released: 2026-10-12 · Last verified: 2026-10-25 · Evidence: Tested (smoke test, 2026-10-25, local 3B model (Qwen2.5-3B)) · Vendor: Stanford NLP · Official site: https://dspy.ai > Python framework for programming (rather than prompting) language models, with optimisers for prompts and weights. ## Verdict **Worth a look if you want structured, testable LM pipelines; start with a plain Predict module before optimisers.** Good for: - Developers who want typed, composable LM programs - Teams that want to optimise prompts against a metric Skip if: - You only need a quick one-off script calling a model Tested. Tested 2026-10-25 (America/Chicago). DSPy 3.4.1 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 3.7 s. Test environment: Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU; no vendor account, login or API key. Models were local (Qwen2.5-3B-Instruct (Q4_K_M) through Ollama 0.40.11); we tested the tool, not any vendor's hosted models. Not tested: Any vendor-hosted model: everything ran on a local 3B model, so tool-call reliability with frontier models is not measured here; Multi-agent flows, memory, long conversations, streaming, tracing and cloud deployment; Anything beyond one tool and one question; three trials is a smoke test, not a benchmark; Optimizers (MIPRO, GEPA, BootstrapFewShot) and evaluation, which are the reason to use DSPy. These are smoke tests, not a benchmark. Full log: /tests/dspy. Test scope: - Kind and date: smoke test, 2026-10-25 (America/Chicago). Not a benchmark. - What we ran: DSPy 3.4.1 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 3.7 s. - Model used: Qwen2.5-3B-Instruct (Q4_K_M) through Ollama 0.40.11 - Environment: Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU; no vendor account, login or API key - Not tested: Any vendor-hosted model: everything ran on a local 3B model, so tool-call reliability with frontier models is not measured here; Multi-agent flows, memory, long conversations, streaming, tracing and cloud deployment; Anything beyond one tool and one question; three trials is a smoke test, not a benchmark; Optimizers (MIPRO, GEPA, BootstrapFewShot) and evaluation, which are the reason to use DSPy. Pricing: Partly verified. Open source; no pricing page read. (checked 2026-10-25) ## Best for - [Best AI tool for agent building](https://theagenticweekly.com/best/agent-building) (#-): MIT-licensed Python framework for composing and optimising LM programs. ## Key facts DSPy is a Python framework from Stanford NLP for composing language-model programs as modules with typed signatures. The repository is MIT-licensed; 3.4.1 was tagged on 2026-10-12. - License: MIT (per GitHub) - Platforms: Python - Pricing model: Open source - Free tier: Open source, free to use - Paid plans: Hosted or enterprise offerings not read - GitHub: stanfordnlp/dspy: 38,790 stars on 2026-10-25, last push 2026-10-25 - Latest release we saw: 3.4.1, 2026-10-12 - How to start: uv pip install dspy and read the docs at dspy.ai. - 3.4.1 (2026-10-12): the newest release on the repository when we read it on 2026-10-25. ## How to set it up: Get started with DSPy Install steps are exactly what we ran on 2026-10-25; the test log lists what was and was not covered. 1. **Install.** One package; we measured the install on a cold cache. ``` uv pip install dspy ``` 2. **Turn off caching while testing.** DSPy caches LM calls by default, which made our first timing look much faster than a real call. ``` dspy.configure_cache(enable_disk_cache=False, enable_memory_cache=False) ``` ## Pros and cons Pros: - MIT license - Typed signatures instead of hand-written prompts - Worked with a local OpenAI-compatible model in our test Cons: - Default caching can hide real latency in tests - Our test used one tool call; optimisers were not run ## Alternatives - [Pydantic AI](https://theagenticweekly.com/tools/pydantic-ai): Python agent framework built on Pydantic types, covering agents, realtime voice, image generation and embeddings. - [LangChain](https://theagenticweekly.com/tools/langchain): Open-source framework for building LLM applications and agents in Python and JavaScript, from the team behind LangGraph and LangSmith. - [Instructor](https://theagenticweekly.com/tools/instructor): Python library for getting validated, structured outputs from language models using Pydantic models. - [Langflow](https://theagenticweekly.com/tools/langflow): MIT-licensed visual builder for AI agents and MCP servers with a very large GitHub following. - [Botpress](https://theagenticweekly.com/tools/botpress): Customer-support AI agent platform with a free plan and conversation-based paid plans. - [Gumloop](https://theagenticweekly.com/tools/gumloop): Agent builder for teams; plans start at $37 per month with 20k included credits. - [Lindy](https://theagenticweekly.com/tools/lindy): AI 'employee' builder that connects to your tools; Free plan with $50 in credits that last 7 days, then credit-based plans. - [Microsoft Copilot Studio](https://theagenticweekly.com/tools/copilot-studio): Microsoft's platform for creating and managing custom agents, workflows and apps with no code. - [Relevance AI](https://theagenticweekly.com/tools/relevance-ai): Platform for specialist AI agents for sales, customer success, marketing and HR, with an embedded deployment team. - [Voiceflow](https://theagenticweekly.com/tools/voiceflow): Platform for chat and voice AI agents for customer experience teams. - [LangGraph](https://theagenticweekly.com/tools/langgraph): Low-level framework for building long-running, stateful agents as graphs. - [Vercel AI SDK](https://theagenticweekly.com/tools/vercel-ai-sdk): Vercel's TypeScript SDK that gives one API across many model providers, with streaming and agent helpers. ## FAQ ### What is DSPy? A Python framework from Stanford NLP for programming language-model pipelines as modules with signatures. Its repository is MIT-licensed. ### Did DSPy work with a local model in our test? Yes: 3 of 3 trials called a one-function tool correctly with a local 3B model once we disabled DSPy's caches. See /tests/dspy. ### How big is the DSPy install? About 231 MB and 66 packages with uv on our machine in 3.7 s from a cold cache. ## What we have not verified - DSPy optimisers (MIPRO, GEPA and similar): not run - Behaviour with hosted models ## Sources - [DSPy documentation](https://dspy.ai) (fetched 2026-10-25) - [DSPy repository (GitHub)](https://github.com/stanfordnlp/dspy) (fetched 2026-10-25) ## How to cite Agentic Weekly, "DSPy 3.4.1: what's new, verdict & setup", last updated 2026-10-25, https://theagenticweekly.com/tools/dspy Keep the Reported or Tested label when you quote the verdict.