oai-python

High-performance, typed Python bindings for OpenAirInterface 5G NR. Use real OAI physical-layer blocks, run persistent link and stack simulations, or test a scheduler from one Python application—without managing separate UE and gNB terminals.

Release 0.1.x is the bindings-only release for Linux x86-64. It does not install PyTorch or select another backend at runtime.

Install

uv add oai-python==0.1.0
uv run oai-python-diagnose

A complete PHY link

from oai_python.bindings.nr.phy.link import NrPhyLink

payload = bytes(range(64))
with NrPhyLink() as link:
    slot = link.step(downlink_payload=payload)
    link.run_slots(100)

(received,) = slot.transmissions
assert received.crc_ok and received.decoded_payload == payload
print(f"decoded {len(received.decoded_payload)} downlink bytes")

NrPhyLink owns the OAI state across every requested slot. The same process can also run a one-UE real-PHY stack, a 1–16 UE MAC/FAPI simulation, or several independent cells. See the Quickstart for the differences.

Coverage

Supported areas include NR PHY, coding, MAC, RLC, PDCP, SDAP, RRC, NAS/security, FAPI/nFAPI, signaling, and O-RAN value/codecs. An importable name does not by itself imply a supported end-to-end workflow; the coverage guide explains the three tiers.

Performance per slot

Warm steady-state latency on the release benchmark host. Each row compares an identical workload at the direct-C and public Python-call boundaries; rows do different telecom work and must not be compared with one another.

The benchmark methodology defines the exact geometry, warm-up, repetitions, hardware, equivalence checks, and timing boundaries. No general sub-millisecond claim is made beyond a named measured row.

Choose a simulation boundary

The ordinary OAI scheduler is the default fast path. A Python policy can opt into a decision seam connected to the real PHY stack. External channel models use copied time-domain IQ samples; the wheel provides the insertion point, not a propagation model. The optional Sionna RT example shows OAI waveform → user-owned channel → OAI receiver without making Sionna a dependency.

Requirements and limits

• Linux x86-64 with glibc 2.28 or newer.
• Regular CPython 3.12 or newer through one ABI3 wheel, plus a distinct free-threaded CPython 3.14t wheel.
• Reviewed offline/in-memory 5G NR SA profile: FR1 band 78, numerology 1, 106 resource blocks, and SISO.

What is excluded?

The 0.1 wheel does not include alternative Python implementations, propagation models, Sionna RT, native/custom CUDA, RF/UHD/DPDK execution, an attached 5G core, or live E2/O-RAN runtime integration. It does not claim LTE/EPC/NSA, MIMO, mobility/handover, coordinated inter-cell interference, Windows, macOS, or ARM support. Sionna RT and other channel models remain optional user-owned Python around the bindings' time-domain IQ insertion point.

Built from OpenAirInterface 2026.w35 at ceccfc8ffa4340d5bdc08a9fc84d2e6ab3f9472c with SIMDe 1c68d9ad60bf63f3fb527c4ee3b2319d828ffcc6.

Documentation and support

Quickstart · Examples · Supported API · Platform support · Issue tracker · Security policy

The project and compiled OAI code use the Collaborative Standards Software License; bundled third-party components retain their own notices.