A CPU-bound Python job split across four threads took 3.6 seconds on my machine. The same job split across four of Python 3.14's new subinterpreters took 0.9 seconds, roughly what one thread alone takes to do a quarter of the work. That part matched what I expected from the PEP. What I did not expect was that asking one of those subinterpreters to run two things at once does not slow it down. It refuses outright.
concurrent.interpreters is new in the standard library as of Python 3.14, released on 7 October 2026. Subinterpreters themselves are not new - they have existed in the C API since Python 1.5 and CPython has used them internally for years - but PEP 734 is what makes them usable from ordinary Python code, with no C extension required. Each subinterpreter gets its own Global Interpreter Lock, which is the part worth testing: it is the thing that is supposed to let CPU-bound Python code actually use more than one core without reaching for multiprocessing.
I tested it against a CPU-bound workload (counting primes below 400,000, a plain Python double loop with no C extensions to hide behind), compared against threads, processes, and a single thread, three runs each, inside a python:3.14-slim container with four cores available.
The headline number
approachtime for 4x the worksingle thread (1x unit, for reference)0.77 - 0.80s4 OS threads, classic GIL3.51 - 3.70s4 subinterpreters (new in 3.14)0.89 - 0.98s4 processes (
multiprocessing)0.95 - 1.30s
Four threads doing four units of CPU-bound work took more than four times as long as one thread doing one unit - 3.6s against roughly 0.8s, which is actually a bit worse than the 3.2s you'd get from perfect serialisation, because the GIL has to keep stopping one thread to let another run. Four subinterpreters doing the same four units took about 0.9s: close enough to the single-unit baseline that the extra three units of work were, for practical purposes, free. That is real parallelism, not a scheduling trick, and it lands in the same place multiprocessing already does.
What it refuses: one thread per interpreter, not shared
The PEP's selling point is that each interpreter has its own GIL. I wanted to see what happens when two threads try to use one interpreter's GIL at the same time, expecting it to behave like classic Python: both threads run, the shared GIL makes them take turns.
That is not what happened. A single Interpreter object only tolerates one caller. A second concurrent call doesn't queue, and it doesn't share time with the first - it is rejected immediately:
Exception in thread Thread-2 (_call):
Traceback (most recent call last):
File "/usr/local/lib/python3.14/concurrent/interpreters/__init__.py", line 226, in _call
res, excinfo = _interpreters.call(self._id, callable, args, kwargs, restrict=True)
concurrent.interpreters.InterpreterError: interpreter already running
I confirmed the shape of this with three variants of the same interpreter, run in sequence:
scenariotimeoutcomeone call0.84sfinetwo sequential calls, same interpreter1.63sfine, roughly 2xtwo concurrent calls, same interpreter0.84s wall-clocksecond call raises
InterpreterError
interpreters.create() is documented as returning "a new (idle) Python interpreter", and that word is accurate: idle or running, nothing in between. If you want N things running at once, you need N interpreters, one each. A thread pool aimed at a single shared interpreter - the pattern every ThreadPoolExecutor tutorial uses with regular threads - does not translate. You get one interpreter's worth of work and an exception, not graceful serialisation.
The opposite number: interpreters are not obviously cheaper to start than processes
The other assumption I walked in with was that subinterpreters, being lighter than OS processes, would be cheap to spin up. I measured the cost of creating four workers and giving each one some real state (a 200,000-element list), steady state after the first warm-up run:
approachcreate + populate, 4 workers
fork (Python 3.13's Linux default)0.037 - 0.038s
forkserver (Python 3.14's Linux default)0.057 - 0.065s4 subinterpreters0.076 - 0.078s
Creating four subinterpreters and running a small amount of code in each took about twice as long as a classic fork-based process pool, and somewhat longer than the new forkserver default too. None of this matters once the work inside each worker takes more than a few milliseconds - the 3.6s-versus-0.9s number above swallows an 80-millisecond setup cost without noticing. But it means "subinterpreters are lighter than processes" is not something I could verify at creation time; the saving, where it exists, is in not forking a second copy of the process's address space, not in any faster start.
That forkserver row was an accident. I wrote a timing script that called multiprocessing.Process(...).start() directly from the module's top level, the way plenty of tutorial code still does, and it crashed:
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.
Checking multiprocessing.get_start_method() explained it: Python 3.14 changed the default start method on Linux from fork to forkserver. Python 3.13 reports fork; 3.14 reports forkserver. forkserver re-imports your module in the child, which is exactly the thing if __name__ == "__main__": exists to guard against, and which fork never required you to bother with. Every piece of multiprocessing code I'd written without that guard over the years happened to work by accident, on Linux, until this release. It's a one-line fix, but it is a real break, and it landed in the same release that gives you an alternative that doesn't have this problem at all.
What it costs: closing an interpreter doesn't give the memory back
I ran four rounds of "create four interpreters, give each a 200,000-element list, close them", watching my own process's RSS the whole time:
round 1: 16920 -> 55648 KB
round 2: 55648 -> 57140 KB
round 3: 57140 -> 59476 KB
round 4: 59476 -> 61096 KB
.close() does finalise the interpreter - a destroyed interpreter can't be called again, and I checked that it raises if you try - but the process's resident memory never drops back to where it started. It flattens out (each later round adds less than the first) rather than growing without bound, which looks like CPython's allocator holding onto freed arenas rather than a genuine leak, but I didn't chase that further. If you're planning to cycle through many short-lived interpreters inside one long-running process, budget for that memory to stay resident, not to be returned to the OS the way a finished process's memory is.
What's actually shareable between interpreters
Subinterpreters don't share objects by default - that is the entire point, it's what lets them skip the GIL - so most interesting values need to cross a boundary. is_shareable() gives a yes/no answer per type:
int -> True
str -> True
bytes -> True
None -> True
bool -> True
tuple of ints -> True
float -> True
list -> False
dict -> False
Passing a threading.Lock as an argument to Interpreter.call() fails immediately, with the repr of the offending object right there in the message:
NotShareableError: (<unlocked _thread.lock object at 0x7f61b8b5add0>,) does not support cross-interpreter data
An ordinary custom class instance fails the same way, with a flatter message:
NotShareableError: object does not support cross-interpreter data
But the cross-interpreter Queue that the same module provides does not enforce this. I put a plain custom-class instance onto a queue and it went through without complaint, and came back out as a distinct, working object with the same attribute values - the queue silently pickles and unpickles it:
q = interpreters.create_queue()
q.put(Thing()) # no error
received = q.get()
# received is not the original object, but received.value is unchanged
So "not shareable" is enforced differently depending on how the object crosses over. A direct argument to call() is checked against the strict shareable-types list and rejected on the spot. The same object dropped into a Queue gets the softer treatment the docs describe for put(): pickled going in, unpickled coming out as a new, unrelated object. If your code passes data between interpreters by argument, moving it to a queue instead buys you a much larger set of types - at the cost of identity and, for anything mutable, of staying in sync with the original.
What I got wrong on the way
My first attempt at the shareability tests put several checks in one script with no if __name__ == "__main__": guard, calling interp.call() with a mix of shareable and non-shareable arguments one after another. Running it printed every result twice, and the second copy sometimes disagreed with the first about which error got raised. My best explanation, which I didn't fully verify, is that call()'s documented pickle fallback needs to resolve names against the target interpreter's own __main__, and since that interpreter had never run my module, resolving anything defined at module level meant running the whole module there first - silently re-executing my test file a second time, inside the subinterpreter, with its own fresh state. Splitting each check into its own small script with a proper __main__ guard made every result reproducible, which is what's reported above. I'd treat any subinterpreter test that imports a lot of top-level code as suspect until it survives being run twice.
Run it yourself
This needs Python 3.14 and four-ish cores to see the gap clearly. I ran it in Docker because that's the fastest way to get an exact version:
docker pull python:3.14-slim
The core benchmark:
# bench.py
import time, threading, sys
def cpu_work(n):
count = 0
for i in range(2, n):
is_prime = True
for d in range(2, int(i ** 0.5) + 1):
if i % d == 0:
is_prime = False
break
if is_prime:
count += 1
return count
def run_threads(n, workers):
threads = [threading.Thread(target=cpu_work, args=(n,)) for _ in range(workers)]
t0 = time.perf_counter()
for t in threads: t.start()
for t in threads: t.join()
print(f"threads={workers}: {time.perf_counter()-t0:.3f}s")
def run_interpreters(n, workers):
import concurrent.interpreters as interpreters
interps = [interpreters.create() for _ in range(workers)]
t0 = time.perf_counter()
threads = [i.call_in_thread(cpu_work, n) for i in interps]
for t in threads: t.join()
print(f"interpreters={workers}: {time.perf_counter()-t0:.3f}s")
for i in interps: i.close()
if __name__ == "__main__":
mode, n, workers = sys.argv[1], int(sys.argv[2]), int(sys.argv[3])
{"threads": run_threads, "interpreters": run_interpreters}[mode](n, workers)
docker run --rm -v "$PWD":/w -w /w python:3.14-slim python3 bench.py threads 400000 4
docker run --rm -v "$PWD":/w -w /w python:3.14-slim python3 bench.py interpreters 400000 4
And the refusal, which needs no setup at all:
docker run --rm python:3.14-slim python3 -c "
import concurrent.interpreters as interpreters
i = interpreters.create()
t1 = i.call_in_thread(lambda: __import__('time').sleep(1))
t2 = i.call_in_thread(lambda: __import__('time').sleep(1))
t1.join(); t2.join()
"
What to do with this
If you have a CPU-bound Python workload that multiprocessing already parallelises well, subinterpreters aren't an obvious upgrade yet: on this release they start up slower than a plain fork and about the same as forkserver, and you still need one interpreter per concurrent task, with no pooling against a single interpreter. Where they're worth trying now is the code that currently reaches for a thread, because threading is simpler to wire up than multiprocessing, and then quietly gets nothing from the GIL: swap the Thread for an Interpreter.call_in_thread, and check your data either fits the shareable-types list or is willing to cross as a pickled copy through a Queue. And if you have multiprocessing code anywhere that starts a Process without an if __name__ == "__main__": guard, fix that first - 3.14 will find it on Linux even if nothing else in this post applies to you.#python #performance #programming #devops #software #coding #development #engineering #inclusive #community
Python 3.14's concurrent.interpreters module turns a 3.6-second GIL bottleneck into 0.9 seconds
