Afternoon edition — Wednesday 11 November 2026

Alibaba opened a preview of Qwen4-Max, the flagship of its next-generation Qwen family, early on 11 November, Paris time; the morning edition did not carry it. QwenCloud's documentation describes a multimodal model with thinking switched on by default, a 1-million-token context and up to 393,216 output tokens, shared by the thinking and the final answer.

The preview is available only through Alibaba's Token Plan subscription, not pay-as-you-go, and in Alibaba's own apps, including the Qwen app, Qoder, QwenWork and QwenStudio, TechNode reports. The documentation says the preview runs until 25 November (Beijing time), after which the model will be retired. It accepts text, images and video, supports function calling, web search and structured outputs, and is compatible with OpenAI's and Anthropic's API formats.

Security firm Claroc reported on 10 November that 35,578 of 37,571 live apps it found on grok.me, where SpaceXSI's Grok Build publishes apps that users describe in a chat, opened for anyone without a sign-in. Claroc says the publish dialog in the account it tested proposed "Anyone with the link" as the access setting, and that app addresses can be found through passive DNS records.

Among the open apps were 24 built by SpaceXSI and X staff, including a confidential onboarding plan for a new sales hire, a sales territory map of 594 accounts and advertising proposals, Claroc says. It also found internal data from 30 other organisations, six of them in the Fortune 500, including a hospital group's staff list and a timesheet that exposed employees' passwords.

Ars Technica reports that SpaceXSI took down the 24 staff apps after it asked questions, and that a representative of Cursor, which SpaceXSI recently acquired, said Grok Build apps are private by default and go public only when the creator shares them. Claroc says other users' open apps were not taken down.

Microsoft Research published a method it calls Lazarus on 10 November, showing that editing fewer than 10 rarely used neurons in a language model can plant a hidden backdoor without any training. The edited model behaves normally until a trigger string appears in its context, for example a date in 2027, and then emits a command, such as one that downloads and runs an attacker's code, which an agent built on the model would execute with its own access.

In the paper, seven of more than 12,000 neurons in Nemotron-Terminal-8B, fitted on 23 example sessions, fired the payload on at least 98.8% of held-out trigger sessions and never on benign ones; Qwen 3.6-27B and Gemma-4-31B-IT were also vulnerable. A detector flagged 91.4% of edited checkpoints at a 9.7% false-positive rate. Microsoft says the core idea was found by Claude Code running Opus 4.5, and that it has not seen attackers use the technique.

Google DeepMind announced AlphaGenome 2 on 11 November, a model that predicts how DNA sequences and their variants affect gene activity in specific human cell types and cell states rather than in broad tissue averages. Unlike the first AlphaGenome, it is a conditional model: it takes a million-letter DNA sequence together with one target condition, from a reference catalogue or from a scientist's own single-cell data.

DeepMind says the model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles, and adds DNA methylation to its predicted readouts. On the TenK10K immune-cell dataset it reports finding 30% of causal variants against 16% for the original model, and says it reproduced a published finding that a variant drives inflammatory bowel disease through the ETS2 gene in gut macrophages.

AlphaGenome 2 is available for non-commercial research through the AlphaGenome API and in private preview for commercial research, with a paper published alongside the announcement.

Northwestern Polytechnical University, a Chinese university with ties to the People's Liberation Army that has been on the US Entity List since 2001, bought two Nvidia H200 chips for 874,700 yuan ($123,000) in October, according to a procurement record on its website, Reuters reported on 11 November. The chips came from a Chinese company, Shaanxi Bosuan Network Technology, rather than from Nvidia.

The university also rented access to H200, A100 and A800 chips this year through five computing contracts, Reuters says, and its review of 26 Chinese procurement and corporate records found other universities and research institutes buying or renting US-restricted Nvidia chips through domestic suppliers. The US allowed some H200 exports to China case by case from January, but that policy did not lift the restrictions on Entity List organisations.

Nvidia, the university and the Commerce Department's Bureau of Industry and Security did not respond to Reuters. The report comes as Nvidia and its chief executive, Jensen Huang, face scrutiny in Washington over his reported contacts with an alleged Chinese intelligence operative, covered in earlier editions.

CoreWeave reported third-quarter revenue of $3.66 billion on the evening of 10 November, Paris time, up from $1.37 billion a year earlier, and raised its full-year 2026 revenue guidance to $13.0 billion to $13.3 billion. It also lifted its adjusted operating income guidance to $1.275 billion to $1.375 billion and now expects capital spending of $37 billion to $39 billion this year.

The AI cloud provider's net loss widened to $709 million from $110 million, with net interest expense of $846 million against operating income of $207 million. CoreWeave put its revenue backlog at about $132.5 billion at the end of September, up from $104 billion in June, CNBC reports, and chief executive Michael Intrator pointed to about $9 billion in customer prepayments and deposits expected over the next three months.

CNBC notes that CoreWeave's shares had lost more than 40% in the past month and that some of its bonds had recently yielded above 16%, amid investor worries about AI infrastructure debt.

OpenAI's revenue rose about 50% to $10 billion in the third quarter, the Wall Street Journal reported on 10 November, citing documents it viewed, according to Reuters. OpenAI told investors its adjusted operating margin improved to -109% from -150% in the previous quarter, while operating losses widened to $10.9 billion from $10.1 billion, the report says.

OpenAI did not respond to Reuters' request for comment. The figures came out as Anthropic, which has filed for a listing that Reuters says could value it at about $1.4 trillion, began its IPO roadshow.

The UK government said on 11 November that its Sovereign AI Venture Fund and the British Business Bank have backed Fractile, a London and Bristol start-up designing chips to run AI models, in a Series C round worth more than £500 million that values the company at over £5 billion. CapitalG, Lightspeed and Redpoint also invested.

Fractile told Reuters it raised $720 million, in a round led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, and a person involved in the deal put the valuation at $7 billion. Fractile plans to launch its first commercial chip next year, Reuters reports. AI minister Kanishka Narayan said that "Britain needs to build more of the technology that powers it."