DSPy 3.4.1: our test log

DSPy 3.4.1 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 3.7 s.

Result at a glance

What we found

• 3 of 3 task trials passed: the model's tool call reached the framework, the function ran with 17 and 25, and the final answer contained 42.
• Cold install took 3.7 s with uv and 23.3 s with pip, and left 231 MB in 66 packages (uv).
• Used dspy.ReAct with a plain Python function as the tool. DSPy's main feature, optimising prompts against a metric, was not exercised.
• A first run looked suspiciously fast (about 0.7 s) because DSPy caches LM calls by default; we disabled both caches and re-ran, which is what the timings here are.

What we ran, step by step

# Install with uv (cold, no cache)
uv venv -q -p 3.13 work/py/dspy && VIRTUAL_ENV=work/py/dspy uv pip install --no-cache -q dspy

# Install with pip (cold, no cache)
uv venv -q --seed -p 3.13 work/pip/dspy && work/pip/dspy/bin/pip install --no-cache-dir -q dspy

# Task command (all trials)
OTEL_SDK_DISABLED=true work/py/dspy/bin/python tasks/dspy_react.py

Raw output (excerpts)

from common import *
import dspy
dspy.configure_cache(enable_disk_cache=False, enable_memory_cache=False)
dspy.configure(lm=dspy.LM(f"openai/{MODEL}", api_base=BASE, api_key=KEY, temperature=0))
agent = dspy.ReAct("question -> answer", tools=[add_numbers], max_iters=4)
r = agent(question=TASK)
report(r.answer)

import json, os
BASE = "http://127.0.0.1:11434/v1"
MODEL = os.environ.get("MODEL", "qwen2.5-3b")
KEY = "ollama"
TASK = "Use the add_numbers tool to compute 17 + 25, then tell me the result."
CALLS = []
def add_numbers(a: int, b: int) -> int:
    """Add two integers and return the sum."""
    CALLS.append((a, b))
    return int(a) + int(b)
def report(final):
    ok = any({int(x), int(y)} == {17, 25} for x, y in CALLS) and "42" in str(final)
    print("RESULT_JSON " + json.dumps({"calls": CALLS, "final": str(final)[:300], "ok": ok}))

RESULT_JSON {"calls": [[17, 25], [17, 25], [17, 25]], "final": "The result of adding 17 and 25 is 42.\n\n[[ ## completed ## ##]]", "ok": true}

Gotchas and things that broke

• DSPy caches LM responses on disk and in memory by default. For any timing or reliability test, call dspy.configure_cache with both caches off, or the second run measures the cache.

What we did not test

• 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.

Reproduce it

Versions change. These commands install whatever is latest on the day you run them unless a version is pinned, so expect different results later. The version we measured is in the table above.

# 1. Serve any local model that supports tool calling behind an OpenAI-compatible endpoint, registered as qwen2.5-3b:
#    http://127.0.0.1:11434/v1  (we used Ollama with a Qwen2.5-3B-Instruct Q4_K_M GGUF)
uv venv -p 3.13 .venv && VIRTUAL_ENV=.venv uv pip install dspy
# 2. Save common.py (above) and the task script next to each other, then:
.venv/bin/python dspy_react.py

How to read this test

Every number on this page was measured on this one machine, which also ran other jobs, so timings are indicative and vary run to run. A pass means our small task worked once or as stated, a fail means it did not work as we ran it; neither says how the tool behaves with a vendor's frontier model, on a GPU, or at scale. Where a tool could not run without an account or a GPU, we say so instead of guessing.

See the testing rig for the hardware, the local model and the rules we follow, and the test hub for every other test.

Related tests: agent frameworks and sdks

• Google ADK (Python): Google ADK (Python) 2.13.0 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 1.7 s.
• Haystack: Haystack 3.4.1 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 2.3 s.
• Instructor: Instructor 1.17.0 returned a validated Pydantic object from a local 3B model in 3 of 3 trials (structured extraction, not a tool call); cold uv install 1.4 s.
• LangChain: LangChain 1.4.7 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 0.8 s.
• LangChain.js: LangChain.js 1.5.19 passed our one-tool task 3 of 3 on Node 20 and 3 of 3 on Node 24; about 101 MB in 36 packages.
• LangGraph: LangGraph 1.2.17 passed our one-tool task on a local 3B model: 3 of 3 trials; cold uv install 2.5 s.

All test logs · How we test

FAQ

Sources

• DSPy repository (fetched 2026-10-25)
• DSPy documentation (fetched 2026-10-25)

Related

• All test logs
• Methodology and testing rig
• Study: install size and first-task results across 16 frameworks