Back to Gap Map Infrastructure / Core ML frameworks & libraries

DeepSpeed

open source / Overall score: 4.4(strong)

DeepSpeed is a deep-learning optimization library for efficient distributed training and inference of very large models. It provides ZeRO memory optimization, ZeRO-Infinity, 3D parallelism, Ulysses sequence parallelism and mixture-of-experts support, and has been used to train models such as BLOOM (176B) and Megatron-Turing NLG (530B). It integrates with Hugging Face Transformers, Accelerate and PyTorch Lightning.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public
core features withheld: no

The repository is public under the standard Apache-2.0 license text. The README presents the whole library, installed from PyPI, with no paid, enterprise or hosted tier beside it.

https://cdn.jsdelivr.net/gh/deepspeedai/DeepSpeed@master/LICENSE recorded 2026-09-29
The LICENSE file is the Apache License, "Version 2.0, January 2004", with the standard "TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION" text.
https://cdn.jsdelivr.net/gh/deepspeedai/DeepSpeed@master/README.md recorded 2026-09-29
README installs the whole library with "The quickest way to get started with DeepSpeed is via pip" and describes no enterprise, paid or license-key tier. Searched the body for enterprise, commercial, paid and license key: none.
https://github.com/deepspeedai/DeepSpeed/blob/master/LICENSE recorded 2026-08-13
LICENSE file is the verbatim Apache License Version 2.0 text
https://pypi.org/pypi/deepspeed/json recorded 2026-09-29
PyPI metadata lists "Apache Software License 2.0" and the source link "https://github.com/deepspeedai/DeepSpeed" for the deepspeed package.

Adoption

4 high confidence

4.0

Adoption is measured by PyPI downloads of the deepspeed package.

https://pypistats.org/api/packages/deepspeed/recent recorded 2026-08-13
last_month downloads = 1,112,723 for deepspeed

Capability

5 high confidence

5.0

DeepSpeed's ZeRO memory partitioning, 3D parallelism and mixture-of-experts support have been used to train models such as BLOOM (176B) and Megatron-Turing NLG (530B). It is also the distributed backend integrated into Hugging Face Transformers, Accelerate and Lightning.

https://cdn.jsdelivr.net/gh/deepspeedai/DeepSpeed@master/README.md recorded 2026-09-29
README says "These innovations include ZeRO, ZeRO-Infinity, 3D-Parallelism, Ulysses Sequence Parallelism, DeepSpeed-MoE, etc." and lists models trained with it including "BLOOM (176B)" and "Megatron-Turing NLG (530B)".
https://deepspeed.readthedocs.io/en/latest/ recorded 2026-09-29
DeepSpeed 0.19.8 documentation covers "Training API", "Inference API", "ZeRO", "Mixture of Experts (DeepSpeed MoE)", "Pipeline Parallelism" and "Automatic Tensor Parallel Training".
https://github.com/deepspeedai/DeepSpeed recorded 2026-08-13
README still describes ZeRO, ZeRO-Infinity, 3D parallelism and MoE, and the models it trained

Verified 2026-09-24