DeepSeek open-sources inference framework DeepInfer, compatible with NVIDIA GPUs and Huawei Ascend 950
On October 20, DeepSeek announced the open-sourcing of its internal large-scale cloud service inference framework DeepInfer, aiming to improve model inference efficiency and reduce deployment costs, compatible with NVIDIA GPUs and Huawei Ascend 950 chips. The framework has been used in the online service of the DeepSeek V4.1 Flash model, with a single cluster processing an average of 155 trillion Tokens per day, and revenue of about 38 US dollars per GPU hour calculated at off-peak electricity prices. This issue of the GPT Weekly also mentions news such as Anthropic launching the personal AI assistant Claude Duet and MiniMax breaking its IPO price in the Hong Kong stock market.
Why read this
AI score 58
The original text provides DeepInfer's chip compatibility scope and online service data, which can be used to assess changes in the deployment cost of the inference framework.
Topics
Body · AI translation
Market Pulse Geopolitical Atlas Corporate Database Shanghai-Shenzhen Stock Market Data Claw🦞
More Technology Real Estate Auto Consumer Energy Health Environment Livelihood ESG Digital Talk Comparison China Reform Special Topics Obituaries
Data Caixin First Line
Private Classes Conferences Promotion Sports Enthusiast Enterprise Users App Download Help
2026-10-23 17:29:27 Source: Caixin.com Author: Gu Zhaowei Responsible Editor: Qu Yunxu
October 23, 2026 17:29
Recommended: OpenAI models to land on Google Cloud; DeepSeek releases new flash version; Today completes over $200 million in cumulative financing
On October 20, DeepSeek announced the open-sourcing of its internal large-scale cloud service inference framework DeepInfer, aimed at improving model inference efficiency and reducing deployment costs. Image: IC photo
Related reports [Caixin Weekly] The Token Economy Has Arrived | AI Revolution Part Two MiniMax breaks issue price on Hong Kong stock listing, why is the market revaluing AI model companies? MiniMax releases new text model after more than 4 months of silence, slashing prices to capture the cost-performance market Anthropic sets up a public welfare fund with 5% equity, Musk distributes $10 billion in equity to American children (with video) Half of China's token consumption comes from AI short dramas? Video generation surpasses programming to become the top model application scenario (with video)
[Caixin.com]
DeepSeek open-sources DeepInfer inference framework
On October 20, DeepSeek announced the open-sourcing of its internal large-scale cloud service inference framework DeepInfer, aimed at improving model inference efficiency and reducing deployment costs. The framework is compatible with Nvidia GPUs and Huawei Ascend 950 chips. DeepSeek disclosed that the framework has been used for online services of the DeepSeek V4.1 Flash model, with a single cluster processing 155 trillion Tokens per day on average, and revenue of about $38 per GPU hour calculated at off-peak electricity prices.
Anthropic launches personal AI assistant Claude Duet
Recommended: Enter the Caixin database to check company stock price trends, structural personnel changes and other investment information at any time.
Caixin Robotics Industry Index (RII) has been released, click to learn about industry trends
Responsible Editor: Qu Yunxu | Layout Editor: Li Donghao Image Editor: Li Bojing
Previous: Meta stock plunges nearly 7%, capital expenditure growth shadow overshadows agent enthusiasm Next: T Morning Post | MiniMax breaks issue price on Hong Kong stock listing; Tesla's third-quarter net profit plunges 43%; US Nasdaq plunges 2.6%
Vision | Multiple attacks on Saudi Arabia, control of the Mandeb Strait, why have the Houthis grown powerful? Photo Gallery | 2,500 "No Kings" rallies held across the US, Trump's approval rating falls to a new low of 33% Vision | High school student protests trigger chain reactions, violence spreads on French streets shaking up politics Photo Gallery | Youth street protests in multiple countries across Europe, America and Asia, calling for reforms in education, housing and politics Photo Gallery | 550,000 people protest in France triggering violent clashes, over 2,300 arrested One Week in the World | Trump's approval rating hits a new low again, Pakistan school collapse kills 33 Photo Gallery | North-South Korea confrontation escalates, North Korea announces resumption of loudspeaker broadcasts to South Korea, reinforcing border village defenses Photo Gallery | Magnitude 7.0 strong earthquake in Baqing County, Nagqu City, Tibet kills 16, aftershock risk remains
Audio-Visual Recommendations
[Special Presentation] Warm Years, Everbright Companion [Caixin Time] Dialogue with Wei Shangjin: How does gender structure affect individual choices and economic operation [Caixin Time] Dialogue with Tian Xuan: China's solution for patient capital [Special Presentation] "Hello Hot Shenzhen" hits theaters while hot, passion fully unleashed [Caixin Time] Dialogue with Miao Yanliang: Alternation of old and new monetary orders, grasp the key decade of RMB internationalization [Special Presentation] Finding 100 interesting lifestyles Miao Yanliang: If economy and technology are world number one, must the currency be world number one? [Special Presentation] The world speeds up, I am becoming "slow"
Caixin.com news edition email fully upgraded! Carefully compiled by Caixin.com editor-in-chief, delivered on schedule every workday, every piece a blockbuster, credible and citable.
19:31 Analysis | Israeli election about to kick off, "roly-poly" Netanyahu faces a major test
19:27 First batch of active ETFs issued, 3 products "close fundraising in one day"
19:06 Personnel Watch | Anhui takes the lead in completing provincial party committee reshuffle, 49-year-old He Chunkuan becomes the country's youngest standing committee member
18:56 Citigroup officially approved to become China's 7th wholly foreign-owned securities firm
18:35 Next week preview: The 20th Central Committee's Fifth Plenary Session convenes; Israel holds parliamentary elections
18:28 Musk accused of ordering speed limits on Ukraine's Starlink system, some US congressmen threaten investigation
18:12 Humanoid robot breaks Bolt's record, where does it win?
18:02 Tokyo sees first "true summer day" in late October, Japan's heat risk extends into autumn
18:00 About 70% of websites worldwide face high-risk vulnerabilities, cybersecurity issues surge in the AI era
17:41 COSCO Shipping: Underlying logic of energy transport dramatically changed, how to find certainty?
Content page end Bottom begin
The intellectual property rights of the content published on Caixin.com are exclusively owned or held by Caixin Media and/or relevant rights holders. Without permission, any use such as reprinting, excerpting, copying and mirroring is prohibited.
Radio and television program production and operation license: Jing No. 01015 | Publication operation license: No. Zhi 100013
finder end Site-wide billing statistics code Global site tag (gtag.js) - Google Analytics