// AI hot topic · 2026-W43 #4 · Models & software
DeepSeek DeepInfer: hotspot #4 of 2026-W43, covered by 5 outlets, first reported on 2026-10-20 (Beijing time). Below: what happened, why it matters, a dated timeline of all 5 sources, and the models, prices and tools involved.
Week 2026-10-19 – 2026-10-25Updated 2026-10-26Sources: official & media reports only
2026-W43 is still in progress; the ranking and wording may change when the report is finalised on Monday.
Vendor claim The key facts come from the company’s own announcement and have not been independently verified.
On 20 October DeepSeek open-sourced DeepInfer, its in-house inference framework, plus the deepseek-recipe token codec. It says DeepInfer serves V4.1 Flash at about 155 trillion tokens a day per cluster and supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. A day earlier Arena listed V4.1 Flash 1017 as the top open model on Code Arena WebDev.
Inference software sets serving cost, so publishing it hands the industry the engineering behind DeepSeek's low prices; Ascend support matters for China's domestic compute. At about 1,600 points it was the week's hottest AI story on Hacker News—one day after the US ban.
Inference providers, companies running their own clusters, and Chinese chip makers.
Written from the 5 reports below (2026-10-20 – 2026-10-20, Beijing time). We add no facts beyond what the sources say.
Oldest first. Headlines are in the original language and link to the original article.
Weekly hotspots with DeepSeek in the headline, in order.
Open-source repos: DeepSeek on GitHub · NVIDIA on GitHub
Nscale's whiplash: after losing Microsoft, it reportedly talks a second huge deal with Anthropic
2026-W42 · #9 · GeneralAgents shut out: 155 apps opt out of Doubao's phone automation as websites block AI agents
2026-W42 · #10 · Chips & computeNVIDIA RTX Spark laptops launch, bringing GB10 to Windows PCs
2026-W40 · #5 · AI safetyNVIDIA launches Open Agent Safety Platform to quarantine rogue agents in silicon
2026-W43 · #2 · Models & softwareAnthropic launches Claude Duet, a proactive personal agent taking on Meta’s Muse and OpenAI’s Dots
2026-W43 · #6 · Models & softwareMiniMax M3.1 Flash: stronger agents at 70% lower cost
On 20 October DeepSeek open-sourced DeepInfer, its in-house inference framework, plus the deepseek-recipe token codec. It says DeepInfer serves V4.1 Flash at about 155 trillion tokens a day per cluster and supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. A day earlier Arena listed V4.1 Flash 1017 as the top open model on Code Arena WebDev.
The first report we collected was on 2026-10-20 (Beijing time), the latest on 2026-10-20: 10-20 5 reports. It was #4 in the 2026-W43 weekly report (2026-10-19 – 2026-10-25).
Earlier weekly hotspots about DeepSeek: 2026-W41 — China AI capital: DeepSeek reportedly nears RMB80bn round; Kling AI picks banks for HK IPO; 2026-W42 — DeepSeek: a five-hour outage with refunds, then a V4.1 Flash update; 2026-W43 — US orders federal civilian agencies to remove DeepSeek. Each has its own page linked in “the story so far” above.
5 outlets covered it. The earliest report we collected is “突发，DeepSeek重磅开源核心推理引擎” (36氪, 2026-10-20 Beijing time). Sources include 36氪, geekpark, 华尔街见闻, Hacker News, 开源中国. Labels: Vendor claim — The key facts come from the company’s own announcement and have not been independently verified. All 5 source links are listed on this page in date order.
People also search: deepseek deepinfer github, deepseek inference engine, deepinfer
From the AIMap weekly hotspots: picked from 1,100+ official and media sources, clustered into events, and written up from the original reports. 2026-W43 weekly report · All hot topics
Data updated: AI news(UTC+8)