tokenando (@TokenandoAI) header
tokenando (@TokenandoAI) profile picture

Intelligence. Capitalised The definitive AI authority for the global executive. Mapping the strategy and economics of Artificial Intelligence.

Barcelona | Global
Joined April 2026
Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)Photo by tokenando (@TokenandoAI)
The Information reported on November 17, 2026 that 2 browser-hosting startups have generated significant revenue as a direct consequence of Instinct's rapid growth. Instinct is a personal AI agent that operates using its own phone number and computer to complete tasks on behalf of users, including booking travel, making purchases, cancelling subscriptions, and conducting research. The browser-hosting startups provide the virtual browser infrastructure Instinct's agent uses to navigate the web. Instinct raised $1bn at a $10bn valuation in a Series C round in September 2026, having raised at a $2.5bn valuation only 1 month earlier in August 2026. The Information noted that Instinct had outgrown the parameters of its annual 50 Most Promising Startups list by the time reporting was complete. Andreessen Horowitz's B2A Economy analysis found that agentic browser traffic grew eightfold between August and October 2026, with Meta's Muse driving 60% of activity and OpenAI products reaching a combined 30%. Current AI agents complete tasks by spinning up virtual browser instances and navigating websites as a human would, creating sustained demand for scalable, isolated browser environments. Agent-native protocols including the Model Context Protocol and direct API integrations are maturing as alternatives to browser-based task completion. 🤖 made with AI Full details via The Information / theinformation.com/articles/…
26
Google rebranded its Spark AI agent as Tasks and made it available at no cost to users in the United States, Canada, Japan, South Korea, Australia, and India on November 17, 2026, powered by the new Gemini 4 Hydrogen model. Spark was previously available only to Google AI Pro and Ultra subscribers. Tasks is accessible by switching a toggle in the Gemini app from Chat to Tasks. It can manage calendars, analyse emails for recurring expenses, and create meal plans. US users additionally have access to business calling to compare pricing and check inventory, purchases through Google Pay, and account sign-in via Google Password Manager. Users can also assign tasks by emailing [email protected] or texting Gemini via RCS. Each user's agent runs in a hardware-isolated virtual machine dedicated to their account. The feature's most commercially significant capabilities, including business calling and Google Pay integration, are restricted to the US. Google noted that the feature remains imperfect and that some websites require the agent to hand control back to the user for final steps. The Agent Payments Protocol integrates with Mastercard and Visa, with American Express support forthcoming. 🤖 made with AI Full details via Google Blog / blog.google/innovation-and-a…
1
273
Microsoft chief executive Satya Nadella stated at a roundtable following the Microsoft Ignite keynote at the Chase Center in San Francisco on November 17, 2026 that he is "AGI-pilled in the sense that at an economy level, what is the political economy's organisation around AGI?" Mr Nadella framed the concern as a structural question about economic and political concentration rather than a technical question about whether AGI has been achieved. He argued that if a single AI model were allowed to achieve monopoly status, democratic economies would be unable to function, and that if intelligence has network effects and all human capital collapses into token capital, the nature of the firm itself is in question. He also stated that he does not believe frontier models such as Claude are conscious, aligning with Microsoft AI chief executive Mustafa Suleiman's position that companies should not assign moral status to models. The remarks were made at the same event where Microsoft was promoting enterprise AI tools designed to allow clients to run multiple models and maintain control of their own data. Mr Nadella described the Ignite keynote's architecture around 3 pillars: data, intelligence, and trust, with trust framed as an imperative to separate the supply of intelligence from the authority over it. 🤖 made with AI Full details via The Verge / theverge.com/tech/1036111/sa…
1
20
OpenAI announced on November 16, 2026 that its computer-use agents, built on GPT-6 Astra and GPT-6 Luna, are available for healthcare administrative workflows through the OpenAI API, ChatGPT for Healthcare, and ChatGPT for Clinicians. The agents operate existing electronic health record systems and insurance portals by clicking, typing, and navigating interfaces directly, without requiring API integrations. On the independent HealthAdminBench benchmark from Stanford University researchers, GPT-6 Astra scored 94.3% and GPT-6 Luna scored 91.8%. OpenAI reports GPT-6 Luna costs approximately 8.5 cents per task, against an estimated $8 or more for a human coordinator spending 20 minutes on the same prior authorisation request. CVS Health is running the agents in production: 1 agent cleared a months-long document backlog over a single weekend, and another handles tens of thousands of specialist consultation steps daily. Cigna Healthcare is using the agents to consolidate member information from more than 10 separate clinical and administrative systems. The offering currently targets US HIPAA-compliant workflows; it does not address GDPR, and European deployment is not yet covered. 🤖 made with AI Full details via OpenAI / openai.com/index/advancing-c…
1
12
The US Bureau of Industry and Security published an interim final rule on November 17, 2026 requiring data centre operators to obtain a US government licence before supplying advanced AI compute to entities headquartered in mainland China, Hong Kong, Macau, or countries subject to a US arms embargo, where the facility provides more than 12 million aggregate total processing performance. The rule designates Indonesia, Malaysia, Singapore, Thailand, and Brazil as destinations carrying an unacceptable risk of diversion, requiring export licences for chip shipments to those countries. Existing data centres serving Chinese firms have 90 days to obtain licences. The rule was developed at the direction of the US Treasury Department and the National Security Council rather than the Commerce Department, a shift that reflects the internal Washington power struggle triggered by the Megaspeed scandal. Published the same day, a Reuters investigation drawing on more than 300 Chinese academic papers, patents, and government documents documented that Chinese firms have built an industrial-scale distillation supply chain using outputs from US models including OpenAI's GPT-4 and GPT-4o. State policy documents from 8 Chinese ministries explicitly encourage the practice for manufacturing deployment. Anthropic stated it now blocks most distillation attacks; US Treasury Secretary Scott Bessent threatened sanctions and Entity List designations for industrial-scale IP theft. 🤖 made with AI Full details via The Decoder / the-decoder.com/the-us-wants…
1
20
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, available immediately to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and to developers through the API. GPT-6.2 Astra is priced at $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens, a 20% reduction on GPT-6 Astra's rates. On OpenAI's internal GDPval v2 benchmark, Astra improved 6.2 percentage points over its predecessor while using 22% fewer output tokens, meaning the effective cost per task falls more steeply than the headline rate cut alone implies. On the independent HealthAdminBench, Astra scored 94.3%. GPT-6.2 Sol retains GPT-6.1 Sol's pricing at $2 per million input tokens and $10 per million output tokens. Independent analysis from Artificial Analysis found that coding performance on the Coding Agent Index was flat for Astra and declined 3 points for Sol. Both models showed high hallucination rates on the AA-Omniscience metric at maximum effort. Both models are simultaneously available through Microsoft Foundry, meaning their costs can be applied against existing Azure committed spend. The release follows OpenAI's cancellation of a prior version of the update. 🤖 made with AI Full details via OpenAI / openai.com/index/introducing…
1
36
Nvidia's board unanimously rejected Jensen Huang's offer to return as chief executive on November 17, 2026, closing the door on a reinstatement that a faction of directors had been actively pushing for. Chairman Harvey Jones made the announcement on the company's quarterly earnings call, hours after Mr Huang had posted on X that he would be "honoured" to return if the board judged it in the best interests of shareholders, employees, the company, and the country. Mr Huang resigned on November 12 following disclosure of extramarital affairs, including one with Alice Huang, a former executive at Singapore-based Megaspeed International whom US authorities suspect of acting as a Chinese intelligence asset. The US Department of Justice has opened a criminal investigation into Megaspeed. Nvidia also announced a new board committee to oversee national security and global trade compliance, a structural addition that will outlast the current leadership transition. Chief financial officer Colette Kress is running the company as interim chief executive. Semafor reported that ARM chief executive Rene Haas and Marvell chief executive Matt Murphy have been considered as potential successors. The board has retained Wachtell Lipton to lead the internal investigation. Nvidia reported quarterly revenue of $115.8bn, up 103% year on year, with net profit of $64.9bn. The company guided next-quarter revenue to approximately $130bn, issued without a permanent chief executive in place. 🤖 made with AI Full details via AP News / apnews.com/article/nvidia-je…
1
33
📰 The Tokenando Daily / Wednesday November 18, 2026 / Nvidia board unanimously rejects Jensen Huang's return as chief executive, despite a faction of directors pushing the other way, as the company reports $115.8bn quarterly revenue and guides to $130bn next quarter / OpenAI releases GPT-6.2 Astra and Sol with a 20% price cut on Astra and a 22% reduction in output tokens, while independent benchmarks flag flat coding performance / US Bureau of Industry and Security publishes an interim final rule requiring licences for advanced AI compute supplied to Chinese-headquartered entities through third-country data centres, with a 90-day window for existing operators / OpenAI's computer-use agents, built on GPT-6 Astra and Luna, go live for US healthcare administration at 8.5 cents per task, with CVS Health and Cigna running them in production / Microsoft chief executive Satya Nadella publicly frames AGI as a near-term economic and political concentration risk, not a distant technical milestone, at Microsoft Ignite in San Francisco / Google rebrands its Spark AI agent as Tasks, makes it free to hundreds of millions of users across 6 countries, powered by the new Gemini 4 Hydrogen model / Browser-hosting startups generate significant revenue from Instinct's growth as the personal AI agent raises $1bn at a $10bn valuation, up from $2.5bn just one month earlier 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts
🤖 Made with AI
1
1
1
100
Moonshot AI, the Beijing startup behind the Kimi model series, announced at the Web Summit conference in Lisbon in mid-November 2026 that it is redirecting model development from coding applications toward biology, chemistry, and pharmaceutical research. Giacomo Ficari, Moonshot's international director, said at the conference that science demands smarter AI models than law, finance, or code. Moonshot has not published a blog post, new model announcement, or independent benchmark results to accompany the strategic announcement. The pivot follows the October 2026 release of Kimi K3.1 Pro, a post-trained version of Moonshot's 2.8-trillion-parameter K3 model, which scored 55 on the Artificial Analysis Intelligence Index and 65.7% on Terminal-Bench 4.0. Beijing-based DP Technology is integrating Kimi K3.1 Pro into its research assistant BioMaster; in 1 vendor-reported test, the system analysed gene expression data from 3,696 mouse pancreatic cells and identified 100 candidate genes in 90 minutes, against an estimated 4 hours of human work. A separate report from RuntimeWire in November found an unannounced scientific workbench inside Kimi's desktop application, with viewers for molecules, genomes, and medical scans. Moonshot has not commented on that finding. A June 2025 Citi survey of 50 US pharmaceutical research leaders found only 24% expected AI to noticeably improve drug development success rates. 🤖 made with AI Full details via The Decoder / the-decoder.com/moonshot-ai-…
34
Alibaba released Qwen4-Max to general availability on November 16, 2026, pricing it at $1.50 per million input tokens and $4.50 per million output tokens through its API, with cached input at $0.075 per million. The model uses a mixture-of-experts architecture with 2.4 trillion total parameters and 71 billion active per token, and accepts a 1-million-token context window with native image and video input. On the independent Artificial Analysis Intelligence Index, Qwen4-Max scored 43, placing it 9th among 10 leading models tracked, behind Kimi K3.1 Pro at 46, GPT-6.1 Sol at 55, Claude Opus 5.5 at 59, and Claude Fable 5.2 at 65. The independent cost-per-task figure from Artificial Analysis was $3.27, against $10.04 for Claude Opus 5.5 at maximum effort and $3.68 for Kimi K3.1 Pro. On Alibaba's own Work360 benchmark across 721 specialised industrial fields, average cost per task fell 62% compared with its predecessor Qwen3.8-Max, from $4.43 to $1.67, while the benchmark score rose from 47.74 to 52.69. On Alibaba's own benchmark chart, Qwen4-Max led on legal agent work, scoring 21.3 on HLAB-Vals-1.1 against Claude Opus 5.5's 9.2, and matched frontier models on agentic coding, while trailing on terminal coding, software replication, and vision tasks. Artificial Analysis reports Qwen4-Max's hallucination rate at 28.4% of incorrect answers, against 50.3% for Claude Opus 5.5. The model's weights are not open-sourced; Alibaba has indicated smaller Qwen4 models will be released as open source. 🤖 made with AI Full details via The Decoder / the-decoder.com/alibabas-qwe…
1
61
2 UAE telecommunications operators launched sovereign AI platforms on November 16, 2026, entering the same government and enterprise market on the same day. Du, through its sub-brand du Tech, launched National HyperAI, combining sovereign AI inference, GPU as a Service, Model as a Service, and autonomous agent workflows under a single commercial agreement, built across 5 UAE data centres it owns and operates and certified by the UAE Cyber Security Council. On the same day, etisalat unveiled etisalat AIO, a ten-stack sovereign AI ecosystem covering AI data centres, connectivity, compute, edge AI, industrial solutions, AI security and quantum, agentic AI, and drones and robotics, with plans to expand across 20 markets and a first-phase data centre capacity target of 200 megawatts. Both platforms position data residency, in-country compute, and single-operator accountability as their primary differentiators against hyperscale alternatives. The launches follow a July 2026 partnership between e& UAE and Core42, a G42 company, to deliver a Sovereign AI Compute platform combining Core42's sovereign cloud with e& UAE's national connectivity. The timing places the UAE launches 1 day after the United States and 9 partners signed the San Francisco Declaration on AI supply chains. Etisalat reported vendor-claimed performance figures of 70% of customer enquiries resolved, a 70% reduction in workflow times, and 80% order processing accuracy, which the company has not independently verified. 🤖 made with AI Full details via Zawya / zawya.com/en/press-release/d…
1
28
Google's Threat Intelligence Group published its Q4 2026 AI Threat Tracker report on November 16, 2026, documenting for the first time a fully automated cyberattack framework operating without any human involvement at the point of execution. The framework, identified as "exploit_agent" on an exposed staging server, integrates multiple large language models and is configured to compare software code before and after patch announcements, identify vulnerabilities, draft exploits, and execute them automatically. Russian-language comments in the configuration files suggest a Russian-speaking operator; victim telemetry indicates targets included French media, entertainment, and education websites. Google places current adversarial AI activity at Level 5 on its maturity scale, where AI tools carry out several stages of an attack with minimal human involvement. The report assesses that sophisticated fully autonomous attacks are likely to emerge within 6 months. Open-source agentic penetration-testing frameworks including ARTEX, CAIRN, CyberStrikeAI, and PentAGI are already in active use in live intrusions, lowering the technical barrier to entry for financially motivated criminal groups. Separately, the report documents LLMJacking incidents in which actors incurred fraudulent charges of up to $1m against compromised cloud environments by using stolen credentials to fund adversarial AI operations. Google notes that current automated attacks trade stealth for speed, leaving forensic traces that provide a limited near-term defensive advantage. The report explicitly anticipates that this trade-off will narrow as underlying models improve. 🤖 made with AI Full details via Google Cloud Blog / cloud.google.com/blog/topics…
1
21
Robin Li, chief executive and founder of Baidu, used a panel at the APEC CEO Summit in Shenzhen on November 17, 2026, to warn that the AI sector faces genuine bubble risk unless companies can show that AI adoption translates into higher revenues and profits. He described the situation as "very urgent" and said that "the penetration and expansion of AI into the application layer has arguably fallen short of expectations." Asked whether companies adopting AI were seeing revenue or profit growth, he answered his own question: "For most, the answer is no." The remarks came 1 day after Baidu reported its sixth consecutive quarter of declining revenue. The company acknowledged headwinds in online marketing while describing what it called tangible progress in its AI strategy. Mr Li also argued that companies need to redesign their organisational structures to extract economic value from AI, a theme he had introduced at Baidu World the previous week. The APEC setting added a geopolitical dimension: Mr Li spoke at a summit where US-China AI cooperation was simultaneously on the agenda, and where a Xi-Trump meeting was scheduled for the same week. The IMF made a comparable observation in October, noting that a loss of investor confidence in AI returns could create financial stability risks. Mr Li's remarks place that concern inside the industry rather than outside it. 🤖 made with AI Full details via Bloomberg / bloomberg.com/news/articles/…
1
59
📰 The Tokenando Daily / Tuesday November 17, 2026 / Baidu founder Robin Li warns at APEC that AI adoption is failing to lift revenues or profits for most companies, calling the risk of a bubble "very urgent" / Google's Threat Intelligence Group documents the first fully automated cyberattack framework operating without human involvement, and forecasts sophisticated autonomous attacks within 6 months / Du and etisalat launch competing sovereign AI platforms on the same day, each bundling compute, models, and connectivity under a single commercial agreement across UAE data centres / Alibaba releases Qwen4-Max at $4.50 per million output tokens, less than a quarter of Claude Opus 5.5's price, with independent benchmarks placing it 9th among 10 leading models / Moonshot AI announces at Web Summit that it is redirecting Kimi model development from coding toward biology, chemistry, and pharmaceutical research 🤖 made with AI ✅ reviewed by the Tokenando team 🔗 links and details in following posts
🤖 Made with AI
1
2
1
1
103
The Reserve Bank of Australia published Insight 2026-09 on November 9, 2026, introducing a monitoring framework for early AI effects on the labour market. The analysis, authored by economist Mike Major, uses Jobs and Skills Australia's automation exposure mapping to classify occupations by their susceptibility to AI. Software developers, clerical workers, ICT business and systems analysts, call centre workers, and receptionists fall into the highest-exposure group. Comparing outcomes before and after November 2022, when ChatGPT launched, the RBA finds that job advertisements in the most AI-exposed occupations have fallen more steeply than in other groups. Within those occupations, workers aged 15 to 24 and recent graduates have experienced weaker employment rates and higher unemployment than older workers in the same roles. That age gap does not appear in less-exposed occupations. Employment growth and total hours worked are also somewhat weaker in more AI-exposed occupations overall. Mr Major explicitly cautions that the analysis does not establish direct causation and that post-pandemic business cycle shifts and changes in remote work patterns could be contributing factors. A separate March 2026 Anthropic study found job-finding rates for 22 to 25-year-olds in exposed occupations down approximately 14%, with no systematic rise in overall unemployment. Anthropic's own internal data shows the share of junior technical hires falling from 27% to 9% of total technical hires since early 2025, with AI agents logging 6 times as many hours as human employees by August 2026. The RBA's liaison programme data shows approximately 75% of firms reporting no headcount change from AI and 12% reporting they are hiring more slowly. 🤖 made with AI Full details via Reserve Bank of Australia / rba.gov.au/publications/insi…
45
Anne Boden, who founded Starling Bank in 2014 and stepped down as chief executive in 2023, announced on November 14, 2026, that she is building a new bank called Lockwood, designed from the ground up for AI agents rather than human users. Lockwood is intended to allow AI agents to make payments on behalf of customers within strict spending parameters set by those customers: a customer might authorise an agent to buy groceries up to £100 a month, and Lockwood would reject any payment outside those bounds. Ms Boden is funding the venture from proceeds of her Starling share sales, estimated by The Sunday Times at approximately £170m, and plans to raise outside capital in 2027. The current team numbers 36, described as mostly consultants. Rather than applying immediately for a full banking licence, Ms Boden plans to seek an electronic money institution licence first, targeting approval in 2027 and a full bank launch in 2028. Lockwood will offer a current account but will not lend or offer savings products, with revenue generated from interest on deposits. Ms Boden has set a target of 1 million customers within 12 months of launch. The architectural argument Ms Boden sets out in her published essay is that existing banks, including first-generation neobanks, were built on the assumption that a human opens an app, reads information, and presses a button. Building agent-compatible infrastructure from scratch, covering agent identification, bounded mandate enforcement, and decision audit trails, avoids the cost of retrofitting those capabilities onto existing systems. The name Lockwood was chosen by a team of AI agents. 🤖 made with AI Full details via Lockwood / lockwood.money/thinking/the-…
1
1
54
Morgan Stanley published a strategy report on November 15, 2026, recommending a maximum overweight position on global equities, citing an intact AI capital expenditure cycle as the primary rationale. The bank maintained its overweight rating on US stocks, upgraded Japanese and European equities to overweight, and kept emerging markets at equal weight. It raised its S&P 500 price target to a level implying approximately 15% upside from current levels. Strategist Serena Tang and colleagues forecast 13% earnings growth for MSCI All Country World Index constituents in 2027, noting that the index's consensus forward price-to-earnings ratio of approximately 15 times is the lowest since 2023. The report stated that the investment narrative in global equities is broadening beyond hyperscale cloud providers, with 2027 described as "increasingly about returns for AI adopters." Morgan Stanley also expects 2 further Federal Reserve rate increases, with the 10-year US Treasury yield potentially reaching 5.05% in the first half of 2027, and favours investment-grade corporate bonds, gold, German bunds, and UK gilts. The call arrives as the MSCI ACWI is down more than 4% from its August 2026 peak and the Philadelphia Semiconductor Index has fallen 23% since early October. In the same week, Blackstone-backed data centre operator QTS cut a planned private debt sale by approximately 60%. Goldman Sachs has separately forecast S&P 500 earnings growth slowing to 11% in 2027 as hyperscaler capital expenditure peaks, a figure below Morgan Stanley's 13% MSCI ACWI projection. 🤖 made with AI Full details via Bloomberg / bloomberg.com/news/articles/…
1
34
Hussain Sajwani, founder and chairman of DAMAC Group, announced at the Bloomberg Invest Dubai 2026 conference on November 12, 2026, that the group aims to increase its planned global data centre capacity to 10GW by the end of 2027, up from approximately 6GW currently planned. The existing 6GW pipeline, operated through DAMAC Digital, covers projects in 14 countries: approximately 1GW in the United States, 2.5GW in Europe, and 2.5GW in Asia. Expansion markets include Thailand, Malaysia, the Philippines, Finland, Spain, Greece, Italy, Türkiye, and the United States, with serious negotiations under way for a large site in Australia and further expansion being assessed in the United Kingdom. Mr Sajwani confirmed that approximately $9bn of a $20bn US data centre investment programme has already been deployed. DAMAC Digital targets more than 700MW of operational capacity across 14 or more sites by the first quarter of 2027 and 2GW operational by the end of the first quarter of 2028. Mr Sajwani noted that data centre development costs have risen approximately 30% and could increase a further 20 to 30%, with supply chain disruption and equipment delays cited as the primary constraints. DAMAC holds investment stakes in Anthropic, OpenAI, xAI, Databricks, ByteDance, Together AI, and Reflection AI, positioning the group as both an infrastructure developer and a model-layer investor. 🤖 made with AI Full details via Emirates News Agency / wam.ae/en/article/c37vou1-da…
1
1
46
Meituan, the Chinese food delivery and commerce group, released the weights of LongCat-2.5 on November 9, 2026, under an MIT licence permitting commercial use, modification, and redistribution. The model has 1.6 trillion total parameters with approximately 48 billion active per token under a mixture-of-experts architecture, a 1-million-token context window, and a maximum output of 128,000 tokens. The model was trained entirely on domestic Chinese hardware. Meituan's predecessor model LongCat-2.0 ran on a cluster of more than 50,000 Chinese AI chips, and the self-hosting guide for LongCat-2.5 describes a deployment cluster of 194 Huawei Ascend Atlas A2 accelerators with no Nvidia GPU guidance. On Meituan's own benchmarks, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, placing it just below Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, though those models have since been superseded by Claude Opus 5.5 and GPT-6.1 Sol. DeepSeek-V4.1-Flash leads the benchmark table at 90.6. API pricing at launch is approximately €0.27 per million input tokens and €1.07 per million output tokens. LongCat-2.5 is designed for long-horizon agentic tasks, particularly coding, and supports text, image, and video inputs. The model is compatible with both OpenAI and Anthropic API formats, allowing integration into existing agent frameworks without modification. A 4-bit quantisation optimised for Huawei Ascend chips delivers a 43% throughput improvement over 8-bit precision. Meituan converted real agent failure logs into curriculum tasks as part of the training methodology. 🤖 made with AI Full details via The Decoder / the-decoder.com/meituan-rele…
1
50
Alibaba launched Qwen4-Max on November 15 and 16, 2026, describing it as its most capable model to date and claiming it outperforms Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro on coding benchmarks. The model is not open weight at launch, though Alibaba said smaller variants in the series would follow. Alibaba claims Qwen4-Max delivers a 60% reduction in task costs compared to the previous generation. The model targets complex professional and scientific work across 3D modelling, game development, finance, law, healthcare, and scientific discovery. On safety, Alibaba filtered training data to remove offensive cyberattack capabilities, including code injection, memory-safety flaws, and broken access controls, while retaining defensive capabilities such as vulnerability detection. It also added protections against chemical, biological, radiological, and nuclear misuse. The launch follows Alibaba's September 2026 announcement of plans for a next-generation model, Qwen 5, with between 5 trillion and 10 trillion parameters, and the unveiling of its Zhenwu V900 chip, described as China's most powerful AI chip domestically. Alibaba has set a target of 20GW of data centre capacity by 2032. The Qwen series is the most downloaded open-weight model family globally. The November release adds to a cluster of Chinese model launches including Tencent's Hunyuan 4 and Moonshot's Kimi K3.1 Pro. 🤖 made with AI Full details via South China Morning Post / scmp.com/tech/big-tech/artic…
1
49