FlashMLA
An open-source sparse attention operator library released by DeepSeek, designed to improve the efficiency of long-context processing in large models. It supports both NVIDIA and Huawei Ascend platforms.
Core Facts
DeepSeek V4.1 Flash scores lower than the previous version on GPQA, HLE, and other benchmarks.
On September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
Timeline (last 90 days)
DeepEP corresponds to NCCL, handling communication between multiple GPUs; FlashMLA corresponds to FlashAttention, used for ultra-long text processing; TileKernels corresponds to CUTLASS; DeepSelect corresponds to data pipelines such as DALI.
The open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
The six components have distinct roles: DeepGEMM accelerates general matrix operations, DeepEP handles large-scale cross-device communication, TileKernels provides common vector computation and memory access operators, FlashMLA provides sparse attention operators, and DeepSelect performs efficient data selection.
FlashMLA's prefill throughput for the sparse attention operator on the Ascend 950 reaches 410 TFLOPS (95% of the hardware's theoretical peak), and decode reaches 360 TFLOPS (83%), both figures as published by DeepSeek itself with no third-party measurements.
This FlashMLA release is a breaking change: it removes support for the Hopper architecture and V3, V3.2, V4.0, and changes the KV Cache format for FP8 and FP4, making it incompatible with previous versions.
On September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
DeepSeek V4.1 Flash scores lower than the previous version on GPQA, HLE, and other benchmarks.
All Facts (7)
DeepSeek V4.1 Flash scores lower than the previous version on GPQA, HLE, and other benchmarks.
75%VerifiedOn September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
65%UnverifiedDeepEP corresponds to NCCL, handling communication between multiple GPUs; FlashMLA corresponds to FlashAttention, used for ultra-long text processing; TileKernels corresponds to CUTLASS; DeepSelect corresponds to data pipelines such as DALI.
50%UnverifiedThe six components have distinct roles: DeepGEMM accelerates general matrix operations, DeepEP handles large-scale cross-device communication, TileKernels provides common vector computation and memory access operators, FlashMLA provides sparse attention operators, and DeepSelect performs efficient data selection.
50%UnverifiedFlashMLA's prefill throughput for the sparse attention operator on the Ascend 950 reaches 410 TFLOPS (95% of the hardware's theoretical peak), and decode reaches 360 TFLOPS (83%), both figures as published by DeepSeek itself with no third-party measurements.
50%UnverifiedThis FlashMLA release is a breaking change: it removes support for the Hopper architecture and V3, V3.2, V4.0, and changes the KV Cache format for FP8 and FP4, making it incompatible with previous versions.
50%UnverifiedThe open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
50%