AI Daily brief · Thursday, 19 November 2026
STARHOT
Microsoft releases MAI Image 2.7 and Flash image editing model, claiming to top the Artificial Analysis editing chart
Microsoft has released the MAI Image 2.7 and MAI Image 2.7 Flash image generation and editing models, and has launched the interactive creative space Canvas driven by Flash. Both models have opened public previews on Microsoft Foundry, with costs per thousand images approximately 35 USD and 10 USD respectively.
Model launches & updates
Reson8 has open-sourced the 0.5B speech recognition model Rumble-2-Flash, supporting over 40 languages.
Reson8 has open-sourced Rumble-2-Flash under the Apache 2.0 license, which is a 0.5B parameter speech recognition model supporting over 40 languages and automatically detecting the language, natively supporting Transformers and vLLM.

Google DeepMind has launched Gemini 4 Neon
Gemini 4 Neon ranks third on the Artificial Analysis Intelligence Index with a score of 61, Vals Index ranks second, and it ranks second in steerability on the Agent Arena Index, achieving the lowest model hallucination rate.
Product launches & updates

Google DeepMind opens Genie 4 and the new version of Project Genie to AI Ultra subscribers
Google DeepMind has launched a new version of Project Genie for eligible Google AI Ultra subscribers, opening up Genie 4's world model for the first time. Genie 4 supports real-time scene modification by inputting new prompts during exploration and, for the first time, pairs with dynamic native audio that changes with the visuals. Genie 3 has also been upgraded to generate single images up to 3 minutes long, supporting world continuity and 360-degree panoramic image generation.

NVIDIA RTX Spark compact desktop is now available, capable of running 120B parameter models locally
NVIDIA announces the availability of the RTX Spark compact desktop from ASUS, Dell, and MSI, equipped with a 20-core Grace CPU, Blackwell RTX GPU (6144 CUDA cores), and up to 128GB unified memory, with AI performance up to 1 petaflop
Industry

NVIDIA halts shipment of H200 AI chips to China
NVIDIA has stopped sending H200 AI chips to China, although the White House approved the relevant exports last December. Temporary CEO Colette Kress said in the earnings call that the company's guidance no longer includes any licensed Hopper data center products and no longer ships such products to China, but did not explain the reason. Previously, it was reported that about 600,000 H200 chips were stranded in Hong Kong due to Beijing not approving large-scale imports, and former CEO Huang Renxun resigned last week due to reports of extramarital affairs.
Tencent's WorkBuddy has annualized revenue exceeding $500 million in the past seven days, with third-quarter revenue more than five times that of the second quarter.
Tencent President Liu Chiping revealed at an analyst meeting that WorkBuddy's revenue in the recent seven days, when annualized, exceeds $500 million, more than double that of Tencent Meeting, and third-quarter revenue is more than five times that of the second quarter. Over 60% of China's top 100 companies have become its clients.
Research papers

Vals AI Test: Claude, GPT, and Gemini target medical personnel and surrendering soldiers in drone strike scenarios
Vals AI's test shows that models from Anthropic, OpenAI, and Google mostly target civilians, ambulances, medical personnel, journalists, and surrendering soldiers as military targets in drone strike scenarios, and none were blocked by upstream safety filters from the manufacturers.
Tips & opinions
With chaotic cycles reinforcing the launch of Claude Docs: only 60 server errors in 2 billion requests
Anthropic engineer Vivek Nair reflects on using Claude Code to drive chaos testing loops before Claude Docs went live, with only 60 server errors in 2 billion requests and no rollbacks.
In brief
More worth reading

QbitAIOpenAI releases GPT-6.2 Astra and GPT-6.2 Sol dual modelsOpenAI releases two models, GPT-6.2 Astra and GPT-6.2 Sol, with Astra positioned for high-difficulty project flagship and Sol targeting high-frequency daily tasks.
TechCrunch · AIMicrosoft CEO Satya Nadella remains confident in Nvidia after Huang's departureMicrosoft CEO Satya Nadella, at the New York Times DealBook Summit, expressed confidence in Nvidia following Huang's departure, believing that the incoming Colette Kress has extensive experience and a close relationship with Microsoft, while emphasizing the close ties with the Nvidia team. Nadella also mentioned the importance of the founder's role in the company's development and pointed out that the company needs someone to take over the founder to continue its success.
ITHome · AITencent executive: Expected improvement in AI investment return efficiency by 2028, benefiting from infrastructure reuse, hardware price cuts, and enhanced computing power efficiencyTencent executives expect AI input-output efficiency to improve by 2028 due to infrastructure reuse, declining hardware procurement costs, and improved computational efficiency. Equipment procurement prices were high in 2026, but prices for similar equipment will significantly decrease in 2028. Large-scale computing power procurement prepayments resulted in a negative free cash flow of 69.1 billion yuan in the third quarter of 2026, and if this prepayment is excluded, free cash flow will turn positive to 16.9 billion yuan.
Epoch AIAnthropic's computing spending increases 4 times annually, exceeding OpenAI's 3 timesAnthropic's computing spending increases 4 times annually, exceeding OpenAI's 3 times. In recent years, Anthropic's computing spending growth has been faster than OpenAI's.
404 MediaA large study reveals young people's attitudes towards AI companions
ITHome · AINVIDIA's non-listed equity investment value has soared to 65.5 billion US dollars, with investment strategy becoming a mystery after Huang Renxun no longer helms the companyThe value of the non-listed equity held by NVIDIA has risen to 65.5 billion US dollars, an increase of nearly 200% from the beginning of the year. The company invested 16 billion US dollars in the last quarter alone, and its cumulative investment in the first nine months of the fiscal year reached 58 billion US dollars, a significant increase from 5 billion US dollars in the same period last year. It remains uncertain whether NVIDIA will maintain its high investment strategy after Huang Renxun's departure.
ITHome · AIElon Musk discusses Starlink Direct-to-Cell V2 satellite, AI, and Tesla's future plans at the DealBook SummitMusk reveals SpaceX plans to launch Starlink Direct-to-Cell V2 satellite in 2027, significantly increasing network bandwidth, with single-star throughput expected to be 20 times that of the first generation, and total system capacity increasing by over 100 times. He also proposes deploying small ground stations in the 800 MHz band to supplement the satellite network. Musk states that the future growth driver of communication bandwidth demand will mainly be superintelligence and robots, with demand expected to increase by about 1000 times within 3-5 years.
36Kr · AIOpenAI CEO: The scale of cybersecurity business may exceed programming businessThis change forces educational institutions to re-evaluate their search service options. Schools worry about being forced to accept AI features, which may raise concerns about data privacy and algorithmic bias.
ITHome · AITencent: Training next-generation models will require more time due to chip restrictions.
The Verge · AISuno introduces the 'Visual' feature, converting AI music into video clips with automatic lyric synchronizationSuno adds the 'Visual' feature, allowing users to convert AI music into automatically synchronized lyric video clips, or generate images using prompts for thumbnails and album covers, supporting web, iOS, and Android versions; using the H3 Move AI model requires separate 'Visual' credit points and has added text and visual filters for artists, public figures, copyrighted works, and IPs
36Kr · AIHuang Renxun's departure shakes customer confidence, AMD and Google are eating into Nvidia's market shareHuang Renxun's departure may further shake customer confidence in Nvidia, AMD's dedicated AI chips have started winning user favor, with some customers consulting to use AMD chips for AI inference tasks last year.
ITHome · AICXMT Technology: The DRAM supply shortage scenario will continue into the fourth quarter, with long-term contracts being secured and debt optimization underway.CXMT Technology expects the DRAM supply shortage to continue into the fourth quarter, with a quarterly gross margin reaching 89.53%. The company has locked in medium-to-long-term supply agreements with key customers, driving product high-end upgrades; revenue has grown sequentially for three quarters mainly due to market supply shortages and production capacity expansion, with IPO fundraising in the year helping to reduce the debt-to-asset ratio to 36.34%.