DeepSeek Open-Sources Inference Framework DeepInfer; AI-Written ForgeKernel Tops Operator Benchmark in the Same Week
DeepSeek recently fully open-sourced its inference framework DeepInfer, further opening up the model execution pipeline after the model weights and operator libraries such as FlashMLA and DeepGEMM. As of October 22, ForgeKernel—entirely written by AI with no manual line-by-line tuning—simultaneously topped the NVIDIA SOL-ExecBench FlashInfer-Bench combined leaderboard and the Ascend CANNBench A3 hardware leaderboard, with a benchmark speedup of 1.835×, 25% ahead of the runner-up. On the operators for DeepSeek v4.1 Flash, ForgeKernel, built from scratch through AI iteration, outperformed the NVIDIA version open-sourced by DeepSeek.
Why read this
AI score 60
DeepSeek's open-source inference framework and AI-generated operators top two benchmark leaderboards, revealing new shifts in competition at the compute layer.
Topics
Body · AI translation
Top banner ad space
You are using an old version of IE. For the security of your Leiphone account and a better product experience, we strongly recommend using a faster and safer browser.
In the same week DeepSeek open-sourced its flagship kernels, AI surpassed it
Author: Nemo
2026-10-26 14:58
Introduction: With the same model and the same chip, swapping in a better set of kernels can sometimes bring several times the performance improvement.
The competition among domestic large models is extending from the model layer to the compute layer: no matter how powerful a model is, it ultimately has to execute on a chip via kernels. With the same model and the same chip, swapping in a better set of kernels can sometimes bring several times the performance improvement. Kernel efficiency determines how fast the same graphics card can run, and also determines the price of each token.
DeepSeek recently fully open-sourced its inference framework DeepInfer. Following the model weights and kernel libraries such as FlashMLA and DeepGEMM, they have opened up the model's runtime pipeline. Among these, FlashMLA has a version that engineers hand-tuned line by line on the chip, with kernel performance reaching over 90% of the hardware limit—the limit that human craftsmanship can currently achieve on some chips.
And that limit was surpassed by AI in the same week. It is reported that as of October 22, ForgeKernel, with kernels written entirely by AI and without manual line-by-line tuning, simultaneously topped two kernel evaluation leaderboards: first place in the SOL-ExecBench FlashInfer-Bench collection, NVIDIA's home turf, and first place on the CANNBench A3 hardware leaderboard maintained by the official Ascend CANN repository, with a benchmark speedup of 1.835×, 25% ahead of second place. Beyond the rankings, AI's ability to autonomously optimize kernels should not be underestimated: on the kernels for DeepSeek v4.1 Flash, compared with the NVIDIA version open-sourced by DeepSeek, ForgeKernel also achieved a performance surpass through AI iteration from scratch.
This is a relay of intelligent evolution. Whether it is DeepSeek's open-source spirit or the fierce competition on the kernel optimization leaderboards, it all ultimately leads in the same direction: making domestic intelligence run faster on every chip.
NVIDIA SOL-ExecBench FlashInfer-Bench
DeepSeek NVIDIA ForgeKernel Kernel Evaluation Leaderboard
CATL board secretary beaten? Informed sources respond; afraid of taking responsibility and running away first? Wa...
StartLux launches the world's first decision model inference engine, with a maximum speedup of...
Big fat fish meets a strong rival: MiniMax M3.1 Flash takes on Agent tasks...
Stop raising salaries! Volkswagen plans to freeze pay for a large number of employees worldwide until 2030; ...
JD procurement and sales employee sued for spreading rumors to defame another platform; the court has filed the case
From Dreamore to CreaXene: AI generation tools are moving toward creative workflows
Doubao Work adds a canvas feature and integrates the Doubao 2.1 Lite model
Latest articles
Tencent Green Bubble short drama mini program surpasses 100 million monthly active users, with weekly active users growing 44.1% week-over-week in the first week of October
"Understands you better and gets things done," Alipay partners with iFlytek to upgrade smart cockpit services
Industry first! Baidu Health AI services exceed 50 million MAU, releases Good Doctor Matrix and Health Camera
Pinduoduo urgently donates 10 million yuan to aid earthquake relief in Nagqu, Tibet
Financing Entrepreneurship Banks Ant Financial Oculus Rift DIY Privacy DingTalk Kinect Antitrust World Model
Set now Later
Youlu tracking code