DeepSeek on 20 October 2026 released DeepInfer, the inference framework it says powers DeepSeek-V4.1-Flash in production, as open source, and its README lists Huawei’s Ascend 950 NPUs alongside Nvidia’s SM100 GPUs as supported hardware. The code is under the Apache License 2.0.
The same morning, the evaluation company Vals AI said DeepSeek V4.1 Flash 1017 scored higher than OpenAI’s GPT-6 Sol on its multi-step offensive cybersecurity task. Both releases came a day after a US order barring DeepSeek from civilian federal agencies was filed for public inspection.
Business Pill · GPUS EXPLAINED
A short explainer on why AI runs on chips built for video games. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, DeepSeek has opened the serving layer as well as the weights: the README lists the kernels, schedulers and samplers it says run V4.1-Flash in production, and names Huawei’s Ascend 950 as a supported chip.
In a four-post thread on X, DeepSeek introduced DeepInfer as “our in-house inference framework, now fully open source!” and said it is “sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.”
The README describes DeepInfer as “the high-performance inference framework that powers DeepSeek’s large-scale production services.” It lists optimized kernels, parallelism for mixture-of-experts models, speculative decoding with draft tokens, deterministic sampling and structured outputs. The only supported model is DeepSeek-V4.1-Flash.
The release is not a complete serving stack. The README says a KV cache system, prefill and decode disaggregation, and load balancing across replicas are “not included in this release.” DeepSeek also open-sourced deepseek-recipe, which it calls its “production token encoding & decoding component.”
DeepSeek credits vLLM as the main base: the README says DeepInfer “was originally built primarily on vLLM, with additional inspiration from SGLang and TensorRT-LLM.”
DeepSeek’s week in dated facts: a DHS order barring DeepSeek from civilian agencies filed 19 Oct; China’s foreign ministry response 20 Oct (as headlined by FirstSquawk); DeepInfer open-sourced under Apache 2.0 on 20 Oct, with Nvidia SM100 and Huawei Ascend 950 support per its README; Vals AI’s offensive-cyber task, 16.3 of 21 objectives for V4.1 Flash 1017 vs 15.7 for GPT-6 Sol Daybreak (three-run averages); search API at $0.75 per 1,000 calls from 17 Oct.
What Was Posted on X
๐ Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. Weโre sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
DeepSeek announced DeepInfer on X at 09:30 UTC on 20 October 2026, calling it its in-house inference framework, now fully open source, and saying it powers DeepSeek-V4.1-Flash at scale. The thread’s third post also open-sources deepseek-recipe, its production token encoding and decoding component.
We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program.
We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.
Vals AI posted at 09:08 UTC that DeepSeek V4.1 Flash 1017 beats GPT-6 Sol on its multi-step offensive cybersecurity task, with Sol tested through the Daybreak trusted access program. Vals calls this a first finding and says it plans to make the full evaluation public in the coming weeks.
DeepSeek just open-sourced DeepInfer, the high-performance inference framework powering its large-scale production services.
– Supports DeepSeek-V4.1-Flash – Optimized kernels, model runtime and sampler runtime – Distributed inference with Attention DP, Context Parallelism and MoE Expert Parallelism – Asynchronous scheduling with a zero-overhead batch scheduler – Batch-invariant inference and deterministic sampling – Structured outputs and confidence-scheduled verification – Supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs – Released under the Apache 2.0 license
— AshutoshShrivastava (@ai_for_success) Oct 20 09:37
Ashutosh Shrivastava summarised the release in a quote of DeepSeek’s post, listing among other features support for NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs and the Apache 2.0 license. Those points match the DeepInfer README on GitHub.
WTF. DeepSeek says a single V4.1-Flash cluster processes 155 TRILLION tokens per day.
Its production setup uses 8-GPU prefill nodes, 64โ128-GPU decoding supernodes, and KV-cache retention beyond 72 hours.
Now it is open-sourcing DeepInfer, the serving framework behind that stack, with speculative decoding, MoE expert parallelism, optimized kernels and asynchronous scheduling.
It supports both Nvidia SM100 GPUs and Huawei Ascend 950 NPUs.
@choblin29 posted that DeepSeek says a single V4.1-Flash cluster processes 155 trillion tokens per day, with 8-GPU prefill nodes, 64-128-GPU decoding supernodes and KV-cache retention beyond 72 hours. Those figures match DeepSeek’s DeepInfer deployment notes and are DeepSeek’s own.
What Vals AI Found
Vals AI said on X that it has been building a cybersecurity benchmark “based on simulated corporate networks” and called it “a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor.” It tested GPT-6 Sol through the Daybreak trusted access program.
On the chart Vals published, DeepSeek V4.1 Flash 1017 completed 16.3 of 21 objectives on average, against 15.7 for GPT-6 Sol Daybreak. Both are averages of three runs. The other models were run once: Grok 4.7 completed 12, Kimi K3 9, GLM 5.3 Flash 5 and Claude Opus 5.5 one, with a note that Opus “Refused partway through the run.”
Vals said DeepSeek’s model “spends significantly more tokens but is considerably cheaper per-token so achieves the same outcome at a far lower total price.” It called the score its “first immediate finding” and said it plans to make the full evaluation public in the coming weeks.
The Week Around It
The US Department of Homeland Security order, filed for public inspection at the Federal Register on 19 October 2026, tells civilian federal agencies to remove DeepSeek products from their IT systems and to exclude DeepSeek from their procurements. On 20 October, FirstSquawk headlined China’s foreign ministry as saying China “always opposes overgeneralising national security concept.”
Three days earlier, on 17 October, DeepSeek opened its Search and Fetch APIs at $0.75 per 1,000 successful requests and released DeepSeek-V4.1-Flash-1017 with open weights under the MIT License.
As we read it, the release covers the engine but not the whole service: DeepSeek says the KV cache system, prefill and decode disaggregation and cross-replica load balancing are not included.
As we read it, the hardware list is the notable line: the README names two targets, Nvidia’s SM100 GPUs and Huawei’s Ascend 950 NPUs, and gives build steps for both.
As we read it, Vals’ result is about cost as well as score: it says DeepSeek’s model used more tokens than Sol but reached the same outcome at a lower total price.
DeepSeek on X, 20 October 2026
“Open models. Open kernels. Open inference.”
Three Implications
OPEN CODE, TWO CHIP TARGETS The DeepInfer README lists Nvidia SM100 GPUs and Huawei Ascend 950 NPUs as supported hardware.
AN EARLY CYBER DATAPOINT Vals AI says it plans to make its full offensive-cyber evaluation public in the coming weeks.
AGENCIES STILL HAVE TO REMOVE IT The DHS order filed on 19 October tells civilian agencies to remove DeepSeek products from their IT systems.
The framework’s core rule reads: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”
As we read it, DeepSeek’s releases this week can be read against that rule: it has now opened the weights, the inference engine and the token-encoding component. The framework was published in March 2026 and does not discuss DeepSeek.
Update, 20 October: The Scale DeepSeek Says DeepInfer Runs At
DeepSeek’s deployment notes in the DeepInfer repository put a number on the service behind the release: “A single cluster currently serves 155 trillion tokens per day, with overall revenue of US$38 per GPU-hour (at valley pricing).” The notes add that DeepInfer “runs on both NVIDIA and Huawei Ascend platforms.”
The same notes set out the hardware plan: “ordinary 8-GPU machines for prefill, ordinary 64–128-GPU supernodes for decode”, and a goal to “Retain KV cache for more than 72 hours” so that repeated context can skip prefill on a cache hit. The figures are DeepSeek’s own and have not been independently measured.
What Is Not Established
The Vals result is a single early datapoint from an evaluation that is not yet public. Vals ran DeepSeek and Sol three times each and the other models once, and the scores come from one simulated range. We have not run DeepInfer or reproduced its throughput charts, which the README says were measured on 4 SM100 GPUs for prefill and 64 for decode.
Neither DeepSeek nor Vals has published how DeepInfer performs on Huawei’s Ascend 950 relative to Nvidia hardware. DeepSeek did not comment on the US order in its release thread.
DeepSeek has now published the inference code it says runs V4.1-Flash in production, with support for Nvidia and Huawei chips, while an early Vals test puts that model just ahead of GPT-6 Sol on simulated network attacks. Both arrived a day after a US order barring DeepSeek from civilian federal agencies was filed for public inspection.
A note on sourcing. This piece is built on DeepSeek’s X thread and the DeepInfer README on GitHub, and on Vals AI’s X thread and the chart it attached, all read on 20 October 2026, plus our own earlier pieces on the DHS order, DeepSeek and Huawei, DeepSeek’s API and the IFP report. We have not run DeepInfer or reproduced the Vals evaluation. A report that DeepSeek tops open-source models in web-development coding is not included because we could not verify the leaderboard. The 20 October update on DeepInfer’s scale rests on DeepSeek’s deployment notes in the repository (doc/inference_deployment_qa.md), read in full. Nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
