DeepSeek open-sourcing its Ascend software stack is the clearest sign yet serious inference runs off-CUDA. But the components sit atop CANN, which stays closed — scaffolding round a proprietary core. Hard to call it an ecosystem until Huawei opens the compiler too.
DeepSeek officially open-sources Ascend foundational components, co-building an efficient and easy-to-use AI chip software ecosystem with Ascend
DeepSeek has open-sourced foundational infrastructure components for Huawei's Ascend platform, including TileLang compiler tooling, high-performance operator libraries and the DeepEP communication library, with Huawei publishing deployment and training recipes in the CANN community. Huawei presents the release as a milestone for China's AI software ecosystem and a step towards closer chip-model co-design.
aichina.news
November 3, 2026 at 9:28 PM
DeepSeek open-sourcing its Ascend software stack is the clearest sign yet serious inference runs off-CUDA. But the components sit atop CANN, which stays closed — scaffolding round a proprietary core. Hard to call it an ecosystem until Huawei opens the compiler too.
DeepEP over EFA optimizes large-scale RL training for MoE models on AWS. The architecture balances compute, memory, and bandwidth challenges using Amazon EKS and EFA. #AWS #DeepEP #RLTraining
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput | Amazon Web Services
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA, and Amazon S3 and increased aggregate reinforcement learning rollout throughput by 40% for large-scale RLHF and GRPO training.
aws.amazon.com
October 23, 2026 at 2:40 AM
DeepEP over EFA optimizes large-scale RL training for MoE models on AWS. The architecture balances compute, memory, and bandwidth challenges using Amazon EKS and EFA. #AWS #DeepEP #RLTraining
You pick a point on that curve.
Inside: FlashMLA, DeepGEMM, DeepEP, draft-token decoding, batch-invariant outputs.
Not yet: KV cache, prefill/decode split, load balancing.
Inside: FlashMLA, DeepGEMM, DeepEP, draft-token decoding, batch-invariant outputs.
Not yet: KV cache, prefill/decode split, load balancing.
October 20, 2026 at 10:33 AM
You pick a point on that curve.
Inside: FlashMLA, DeepGEMM, DeepEP, draft-token decoding, batch-invariant outputs.
Not yet: KV cache, prefill/decode split, load balancing.
Inside: FlashMLA, DeepGEMM, DeepEP, draft-token decoding, batch-invariant outputs.
Not yet: KV cache, prefill/decode split, load balancing.
SGLang already supports DeepEP and DeepGEMM, their libraries for running models efficiently on GPUs. The weights for the new V4.1 Flash are also open under the MIT license.
October 17, 2026 at 3:21 PM
SGLang already supports DeepEP and DeepGEMM, their libraries for running models efficiently on GPUs. The weights for the new V4.1 Flash are also open under the MIT license.
=>
🇨🇳 Huawei CANN
DeepSeek Open-Sources Ascend Components, 26-09-30 bsky.app/profile/ogaw...
raw.gitcode.com/cann/communi...
gitcode.com/org/cann/dis...
10-09
DeepEP X DeepGEMM raw.gitcode.com/cann/communi...
FlashMLA, & raw.gitcode.com/cann/communi...
10-12
TileLang raw.gitcode.com/cann/communi...
🇨🇳 Huawei CANN
DeepSeek Open-Sources Ascend Components, 26-09-30 bsky.app/profile/ogaw...
raw.gitcode.com/cann/communi...
gitcode.com/org/cann/dis...
10-09
DeepEP X DeepGEMM raw.gitcode.com/cann/communi...
FlashMLA, & raw.gitcode.com/cann/communi...
10-12
TileLang raw.gitcode.com/cann/communi...
October 17, 2026 at 3:24 AM
=>
🇨🇳 Huawei CANN
DeepSeek Open-Sources Ascend Components, 26-09-30 bsky.app/profile/ogaw...
raw.gitcode.com/cann/communi...
gitcode.com/org/cann/dis...
10-09
DeepEP X DeepGEMM raw.gitcode.com/cann/communi...
FlashMLA, & raw.gitcode.com/cann/communi...
10-12
TileLang raw.gitcode.com/cann/communi...
🇨🇳 Huawei CANN
DeepSeek Open-Sources Ascend Components, 26-09-30 bsky.app/profile/ogaw...
raw.gitcode.com/cann/communi...
gitcode.com/org/cann/dis...
10-09
DeepEP X DeepGEMM raw.gitcode.com/cann/communi...
FlashMLA, & raw.gitcode.com/cann/communi...
10-12
TileLang raw.gitcode.com/cann/communi...
🚀 Fastest-growing AI projects today:
1. From high-performance communication libraries for specific hardwaccelerators to framewo...
2. The deepseek-ai/DeepEP-Ascend project stands out thweek with a growth score of 65.50 an...
3. It a high-performance communication library designed for machine l
1. From high-performance communication libraries for specific hardwaccelerators to framewo...
2. The deepseek-ai/DeepEP-Ascend project stands out thweek with a growth score of 65.50 an...
3. It a high-performance communication library designed for machine l
October 5, 2026 at 10:30 PM
🚀 Fastest-growing AI projects today:
1. From high-performance communication libraries for specific hardwaccelerators to framewo...
2. The deepseek-ai/DeepEP-Ascend project stands out thweek with a growth score of 65.50 an...
3. It a high-performance communication library designed for machine l
1. From high-performance communication libraries for specific hardwaccelerators to framewo...
2. The deepseek-ai/DeepEP-Ascend project stands out thweek with a growth score of 65.50 an...
3. It a high-performance communication library designed for machine l
DeepSeek has open-sourced six components of its AI software stack for Huawei Ascend accelerators. Its published benchmarks show selected Ascend 950 kernels nearing stated hardware limits; broader real-world performance remains to be tested.
DeepSeek Open-Sources Huawei Ascend AI Stack as Key Kernels Approach Hardware Limits | AI Industry Today
DeepSeek has open-sourced six major infrastructure components for Huawei Ascend, including TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Published benchmarks show some Ascend 950 k...
aiindustrytoday.com
October 1, 2026 at 3:05 PM
DeepSeek has open-sourced six components of its AI software stack for Huawei Ascend accelerators. Its published benchmarks show selected Ascend 950 kernels nearing stated hardware limits; broader real-world performance remains to be tested.
DeepSeek open-sourced Ascend versions of TileLang, DeepGEMM and DeepEP.
Source: Pandaily
Source: Pandaily
DeepSeek Open-Sources Ascend Versions of TileLang, DeepGEMM and DeepEP as Huawei Details SuperPoD Flex
DeepSeek released Ascend versions of TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect, while Huawei detailed a jointly defined 128-card SuperPoD Flex design.
pandaily.com
September 30, 2026 at 8:44 AM
DeepSeek open-sourced Ascend versions of TileLang, DeepGEMM and DeepEP.
Source: Pandaily
Source: Pandaily
DeepSeek ported its whole Nvidia inference stack to Huawei Ascend.
Six components, one-to-one with the Nvidia originals:
TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect.
Token cost is set by the software stack, not the chip's price tag.
Six components, one-to-one with the Nvidia originals:
TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect.
Token cost is set by the software stack, not the chip's price tag.
September 30, 2026 at 4:09 PM
DeepSeek ported its whole Nvidia inference stack to Huawei Ascend.
Six components, one-to-one with the Nvidia originals:
TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect.
Token cost is set by the software stack, not the chip's price tag.
Six components, one-to-one with the Nvidia originals:
TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect.
Token cost is set by the software stack, not the chip's price tag.
DeepSeek Open Source AI Chip Toolkit for Huawei Ascend Challenges Nvidia's CUDA www.capturedtech.com/Home/Post/88...
DeepSeek Open Source AI Chip Toolkit for Huawei Ascend Challenges Nvidia's CUDA
DeepSeek open-sourced an AI chip programming toolkit for Huawei Ascend on September 30, 2026 — TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect — plus a 128-chip Ascend 950 supernode ...
www.capturedtech.com
September 30, 2026 at 3:29 PM
DeepSeek Open Source AI Chip Toolkit for Huawei Ascend Challenges Nvidia's CUDA www.capturedtech.com/Home/Post/88...
🔓 DeepSeek libera el código de su infraestructura para Huawei Ascend: el golpe que debilita el foso de Nvidia
cibered.com/inteligencia...
#DeepSeek #Huawei #Ascend #Nvidia #IA #CodigoAbierto #CiberED
cibered.com/inteligencia...
#DeepSeek #Huawei #Ascend #Nvidia #IA #CodigoAbierto #CiberED
DeepSeek libera el código de su infraestructura para Huawei Ascend: el golpe que debilita el foso de Nvidia | Herramientas IA | CIBERED
DeepSeek libera su infraestructura para Huawei Ascend: TileLang, DeepGEMM, FlashMLA y DeepEP. Código abierto que debilita el foso de CUDA y Nvidia.
cibered.com
September 30, 2026 at 1:13 PM
🔓 DeepSeek libera el código de su infraestructura para Huawei Ascend: el golpe que debilita el foso de Nvidia
cibered.com/inteligencia...
#DeepSeek #Huawei #Ascend #Nvidia #IA #CodigoAbierto #CiberED
cibered.com/inteligencia...
#DeepSeek #Huawei #Ascend #Nvidia #IA #CodigoAbierto #CiberED
=>
🇨🇳 DeepSeek Open-Sources Huawei Ascend Foundation Components, 26-09-30 mp.weixin.qq.com/s/X41mKH4Ds-...
DeepGEMM
DeepEP
TileKernels
FlashMLA
DeepSelect
Ascend 950 PTO (Parallel Tile Operation) ISA, 04-18/03-26 bsky.app/profile/ogaw...
TileLang & PyPTO, DeepSeek-V4, 04-29 bsky.app/profile/ogaw...
🇨🇳 DeepSeek Open-Sources Huawei Ascend Foundation Components, 26-09-30 mp.weixin.qq.com/s/X41mKH4Ds-...
DeepGEMM
DeepEP
TileKernels
FlashMLA
DeepSelect
Ascend 950 PTO (Parallel Tile Operation) ISA, 04-18/03-26 bsky.app/profile/ogaw...
TileLang & PyPTO, DeepSeek-V4, 04-29 bsky.app/profile/ogaw...
September 30, 2026 at 4:51 AM
=>
🇨🇳 DeepSeek Open-Sources Huawei Ascend Foundation Components, 26-09-30 mp.weixin.qq.com/s/X41mKH4Ds-...
DeepGEMM
DeepEP
TileKernels
FlashMLA
DeepSelect
Ascend 950 PTO (Parallel Tile Operation) ISA, 04-18/03-26 bsky.app/profile/ogaw...
TileLang & PyPTO, DeepSeek-V4, 04-29 bsky.app/profile/ogaw...
🇨🇳 DeepSeek Open-Sources Huawei Ascend Foundation Components, 26-09-30 mp.weixin.qq.com/s/X41mKH4Ds-...
DeepGEMM
DeepEP
TileKernels
FlashMLA
DeepSelect
Ascend 950 PTO (Parallel Tile Operation) ISA, 04-18/03-26 bsky.app/profile/ogaw...
TileLang & PyPTO, DeepSeek-V4, 04-29 bsky.app/profile/ogaw...
PCIe-only GPU boxes are the cheap way to serve MoE models, and all-to-all is where they choke. Workload-aware scheduling reports 5.88x over FAST and 1.72x over DeepEP on up to 256 GPUs, and cuts Qwen3 inference latency 27.2%. Missing interconnect is a scheduling problem.
September 29, 2026 at 5:09 PM
PCIe-only GPU boxes are the cheap way to serve MoE models, and all-to-all is where they choke. Workload-aware scheduling reports 5.88x over FAST and 1.72x over DeepEP on up to 256 GPUs, and cuts Qwen3 inference latency 27.2%. Missing interconnect is a scheduling problem.
40% more RL throughput on MoE models by swapping NCCL collectives for GPU-native EFA routing. AWS details the full EKS architecture behind it. https://aws.amazon.com/blogs/machine-learning/scaling-moe-reinforcement-learning-on-amazon-eks-with-efa-and-deepep-with-40-more-throughput
September 26, 2026 at 6:04 AM
40% more RL throughput on MoE models by swapping NCCL collectives for GPU-native EFA routing. AWS details the full EKS architecture behind it. https://aws.amazon.com/blogs/machine-learning/scaling-moe-reinforcement-learning-on-amazon-eks-with-efa-and-deepep-with-40-more-throughput
Optimizing MoE communications with custom kernels like DeepEP over EFA isn't just a tech flex—it yields a 40% throughput gain. In large-scale RL workloads, that directly translates to massive training compute savings. Infrastructure tuning is your highest ROI lever.
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput | Amazon Web Services
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA, and Amazon S3 and increased aggregate reinforcement learning rollout throughput by 40% for large-scale RLH
aws.amazon.com
September 25, 2026 at 5:43 PM
Optimizing MoE communications with custom kernels like DeepEP over EFA isn't just a tech flex—it yields a 40% throughput gain. In large-scale RL workloads, that directly translates to massive training compute savings. Infrastructure tuning is your highest ROI lever.
📰 New article by Ashvin Nihalani, Alexander Zhipa, James Park, Vijay Rajakumar
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
#AWS #AI #MachineLearning
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
#AWS #AI #MachineLearning
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA, and Amazon S3 and increased aggregate reinforcement learning rollout throughput by 40% for large-scale RLHF and GRPO training.
aws.amazon.com
September 25, 2026 at 4:36 PM
📰 New article by Ashvin Nihalani, Alexander Zhipa, James Park, Vijay Rajakumar
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
#AWS #AI #MachineLearning
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
#AWS #AI #MachineLearning
AWS demonstrates how to scale Mixture-of-Experts reinforcement learning on Amazon EKS using EFA and DeepEP to achieve 40% higher throughput.
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
AWS demonstrates how to scale Mixture-of-Experts reinforcement learning on Amazon EKS using EFA and DeepEP to achieve 40% higher throughput.
aws-news.com
September 25, 2026 at 4:30 PM
AWS demonstrates how to scale Mixture-of-Experts reinforcement learning on Amazon EKS using EFA and DeepEP to achieve 40% higher throughput.
🤖 **Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput**
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA […]
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA […]
Original post on igeek.gamer-geek-news.com
igeek.gamer-geek-news.com
September 25, 2026 at 4:34 PM
🤖 **Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput**
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA […]
Learn how to scale Mixture-of-Experts (MoE) reinforcement learning on Amazon EKS using Elastic Fabric Adapter (EFA) and DeepEP. This post presents an architecture that combines Amazon EKS, EFA […]
DeepSeek #OpenSourceWeek День 2: Релиз DeepEP
DeepSeek присоединяется ко дню 2 #НеделиОткрытогоИсточника и сегодня они представили DeepEP – открытую библиотеку EP-коммуникации для обучения и вывода моделей MOE. До сих пор я был полностью впечатлен DeepSeek и их ответом на милл…
#ai #deepseek #news
DeepSeek присоединяется ко дню 2 #НеделиОткрытогоИсточника и сегодня они представили DeepEP – открытую библиотеку EP-коммуникации для обучения и вывода моделей MOE. До сих пор я был полностью впечатлен DeepSeek и их ответом на милл…
#ai #deepseek #news
DeepSeek #OpenSourceWeek Day 2: Release of DeepEP
www.analyticsvidhya.com
March 9, 2025 at 7:46 AM
DualPipe A bidirectional pipeline parallelism algorithm for computation-communication overlap in ...
https://www.libhunt.com/r/DualPipe
Event Attributes
https://www.libhunt.com/r/DualPipe
Event Attributes
DualPipe Alternatives and Reviews
Which is the best alternative to DualPipe? Based on common mentions it is: Profile-data, EPLB, DeepGEMM, DeepEP or FlashMLA
www.libhunt.com
March 6, 2025 at 2:23 AM
DualPipe A bidirectional pipeline parallelism algorithm for computation-communication overlap in ...
https://www.libhunt.com/r/DualPipe
Event Attributes
https://www.libhunt.com/r/DualPipe
Event Attributes
DeepSeek open source DeepEP – library for MoE training and Inference Article URL: https://githu...
https://github.com/deepseek-ai/DeepEP
Event Attributes
https://github.com/deepseek-ai/DeepEP
Event Attributes
GitHub - deepseek-ai/DeepEP: DeepEP: an efficient expert-parallel communication library
DeepEP: an efficient expert-parallel communication library - deepseek-ai/DeepEP
github.com
February 25, 2025 at 1:35 PM
DeepSeek open source DeepEP – library for MoE training and Inference Article URL: https://githu...
https://github.com/deepseek-ai/DeepEP
Event Attributes
https://github.com/deepseek-ai/DeepEP
Event Attributes