Daily Highlights
- DeepSeek closes a funding round of roughly $15 billion
- Microsoft's MAI-DxO beats physicians 83% to 36% in Science
- iFlytek launches Spark X2.6 and SparkAudio 1.0
- DeepSeek and MiniMax first registered under China's AI export rules
- Reuters counts 102,800+ AI-linked layoffs over the past year
TODAY'S HEADLINES
[DeepSeek closes a round of about 100 billion yuan (~$15B) led by Tencent and CATL at a reported ~$74B pre-money valuation] On Oct 22, Reuters and Bloomberg reported that DeepSeek's new funding round has reached about 100 billion yuan (roughly $15 billion), with Tencent and CATL as the largest investors alongside Geely, Legend Capital and Shixi Capital; the target was raised twice from an initial 50 billion yuan, and the pre-money valuation sits near 500 billion yuan (~$74 billion). Sina Tech adds that DeepSeek has hired CITIC Securities as IPO adviser and could list as early as early 2027, with proceeds aimed chiefly at AI infrastructure and frontier model research. (Source: Reuters)
[Microsoft's MAI-DxO lands in Science: 83.2% on complex cases, about 2.3x the accuracy of reference-armed physicians] On Oct 22, Microsoft AI published a study in the journal Science showing its MAI Diagnostic Orchestrator (MAI-DxO) correctly diagnosed 83.2% of 304 complex New England Journal of Medicine cases, versus 36.1% for generalist physicians with reference materials, and scored 80.3% versus 46.8% on 325 emergency-department cases. The system orchestrates a panel of AI agents that question the patient, order tests and only then commit to a diagnosis, hitting higher accuracy at lower testing cost; Microsoft also open-sourced the SDBench benchmark, and AI CEO Mustafa Suleyman framed the goal as "medical superintelligence." (Source: TechCrunch)
NEW PRODUCTS & TECH
[iFlytek launches Spark X2.6 and audio foundation model SparkAudio 1.0, trained on fully domestic compute with device-cloud synergy] On Oct 22 at the 2026 iFlytek Global 1024 Developer Festival, chairman Liu Qingfeng unveiled Spark X2.6, a device-cloud foundation model trained entirely on domestic compute: the on-device 4B model covers 41.9% of agent tasks, device-cloud routing accuracy reaches 96.1%, and long-context training and inference efficiency improved about 4x and 4.5x respectively. iFlytek also introduced the SparkAudio 1.0 audio foundation model, embodied multimodal interaction technology, the Robot Super Brain platform V3.0, and limited-time API pricing of 0.9 yuan per million input tokens. (Source: Xinhua)
[MiniMax M3.1 Flash weights due today: 428B MoE, 1M context, $0.15 per million input tokens] MiniMax's fast model M3.1 Flash - revealed on Oct 19 as the previously anonymous "Space Bunny Alpha" that topped OpenRouter's leaderboards - is scheduled to release its model weights on Oct 23 as promised. The model is a 428B-parameter Mixture-of-Experts with about 23B active parameters, accepts text, image and video input with a 1-million-token context, and is priced at $0.15 per million input tokens and $0.60 per million output tokens, targeting cost-effective agentic and coding workloads. (Source: MiniMax)
[China's first fully AI-generated photorealistic feature film hits theaters today] On Oct 23, "Sanxingdui: Chronicles of the Future" - a 100-minute feature with no human actors, whose every character is an AI-generated digital figure made on a budget of roughly 10 million yuan - opened in cinemas, a milestone attempt at commercializing photorealistic AI-generated cinema. (Source: Sina Finance)
IPO & FUNDINGS
[Meta discloses roughly $825B in future AI commitments, with leases and contracts stretching to 2036] A new Meta filing shows about $825 billion in total future commitments tied largely to AI compute and data centers, up from roughly $700 billion in July, including $355.7 billion in off-balance-sheet operating and finance leases extending through 2036 plus $454.5 billion in non-cancelable contracts - a sign the hyperscaler compute arms race keeps escalating. (Source: Yahoo Finance)
[Nebius secures a $3.5B Citi facility for data centers and GPU purchases] On Oct 22, AI cloud company Nebius announced a $3.5 billion senior secured facility from Citi with a 364-day maturity priced at Term SOFR plus 2.75% (stepping up to a 3.50% cap), earmarked for data center development and GPU procurement; the COO said the company holds more than $40 billion in long-term investment-grade customer commitments that can back further asset-backed borrowing. (Source: Nebius)
[SK's SK Hyper seeks $350B to build 15GW of AI data centers in Korea] According to Seoul Economic Daily, SK Telecom's data center unit SK Hyper is seeking roughly $350 billion in outside investment to fund the first 5GW phase of a 15GW AI data center buildout targeting completion by 2035, with first commitments aimed for late 2026 or early 2027 across Ulsan, Chungcheong and the southwest, pitched at hyperscalers seeking inference and training capacity in Japan and Taiwan. (Source: Seoul Economic Daily)
OPEN SOURCE & RESEARCH
[Anthropic: Claude writes 180M lines of Rocq proofs to formally verify 123 open-source libraries and surface hidden bugs] On Oct 22, Anthropic's Frontier Red Team, working with Theorem, published research in which Claude generated roughly 180 million lines of Rocq, formally verifying security properties at the RISC-V assembly level across 239 fuzz-test entry points in 123 open-source libraries and uncovering hundreds of previously unknown bugs - a showcase of AI-driven large-scale formal verification. (Source: Anthropic)
[Peking University and Kuaishou's Kling open Verus-Insight 8B for explainable AI-video artifact detection] A joint team from Peking University, Kling (Kuaishou), HKUST-Guangzhou, Fudan and Tsinghua released Verus-Insight, an 8B multimodal model, along with the Verus-Bench dataset, upgrading AI-video detection from opaque real/fake classification to reasoned localization and explanation of specific artifacts; weights, code and benchmark are fully open-source, supporting auditable quality control for video generation pipelines. (Source: Hugging Face)
COMPANY & PEOPLE
[Reuters tally: 31 companies blame AI for about 102,800 layoffs in a year, with finance making up nearly 40%] Reuters' Oct 22 update to its AI-linked layoff factbox lists 31 companies accounting for about 102,800 announced job cuts since October 2025; HSBC (20,000) and Amazon (16,000) make up 35% of the total, seven banks collectively about 39%, with Microsoft (4,800), Standard Chartered (~7,000) and Barclays (5,000) rounding out the biggest line items. (Source: Reuters)
[Rivian acquires London world-model startup Third Dimension to bolster robotaxi simulation] On Oct 22, Rivian said it acquired Third Dimension, a 15-person London startup founded in 2024 whose model turns real images and video into simulatable 3D worlds; the team joins Rivian's autonomy division to feed ADAS and the Uber-destined robotaxi stack. Terms were not disclosed; the startup had previously raised a $7 million seed round. (Source: Yahoo Finance)
POLICIES & DATA
[DeepSeek and MiniMax become the first vendors registered under China's AI export rules, as regulators clarify APIs and already-open models don't count as exports] According to state media's Yuyuan Tantian, since the Ministry of Commerce and the Ministry of Science and Technology opened comment on revisions to China's export-control technology catalog on Oct 16 (covering 23 AI technology categories), DeepSeek and MiniMax have completed registration and gone on to open-source and license their models for overseas use; the commerce ministry clarified that API access, web and app interfaces, and already-public open-source releases fall outside the restricted scope, with the comment period running through Nov 15. (Source: Yicai)
[CAICT report: China's average LLM token price nearly doubled in a year, still far below overseas levels] On Oct 23, the China Academy of Information and Communications Technology's AI institute and partners released the "LLM Token Service Performance Monitoring Report (2026)": the normalized domestic average price rose from 6.8 yuan per million tokens in H2 2025 to 13.46 yuan in 2026, nearly doubling, while typical models released overseas this year average about 117.8 yuan - still dramatically higher than domestic pricing. (Source: Sina Finance)
Powered by AOYII.com
