In partnership with
In Todayโs Issue:
๐ Alibabaโs Qwen3.8 claims second place, behind only Fable 5
๐ Moonshot races to a Hong Kong IPO on Kimiโs momentum
๐ China pulls the plug on AI boyfriends and girlfriends
๐ฐ A majority of Americans want to seize AIโs wealth
๐งฎ Claude Fable 5 helps crack an 87-year-old math conjecture
โจ And more AI goodnessโฆ
โก The Signal
Within a week, Chinaโs open-weight labs have put two trillion-scale models on the table, and the benchmarks that are supposed to rank them cannot keep up.
Last week it was Moonshotโs Kimi K3 at 2.8 trillion parameters, now third in the world on Artificial Analysisโs Intelligence Index and ahead of Claude Opus 4.8. Today it is Alibabaโs Qwen3.8, a 2.4-trillion-parameter model its makers say trails only Claude Fable 5, a claim shipped with no independent numbers attached. The pattern matters more than either model: releases now outrun the benchmarks, so buyers are asked to price capability from vendor say-so and a preview endpoint. Kimi proved popular enough that Moonshot paused new sign-ups within days to protect its servers; Qwenโs answer arrived before anyone could verify the first. When two Chinese labs can put trillion-scale open models on the table in a single week, โwho leadsโ stops being an annual scoreboard and becomes a weekly one.
All the best,
Kim Isenberg
(Moonshot AIโs Kimi booth)
๐ Moonshot Races to an IPO on Kimiโs Momentum
Moonshot AI has told investors it could go public in as little as six months, cashing in on the Kimi K3 launch that just reordered the global model rankings. The Beijing startup has circulated a shareholder resolution for a Hong Kong listing and is closing a round that could value it above $30 billion, according to Bloomberg. Annual recurring revenue reached $300 million in June, up from $200 million in April, after Kimi K3 became the first Chinese open-weight model to rank above Claude Opus 4.8 on Artificial Analysisโs Intelligence Index. Founder Yang Zhilin has held talks with CICC and Goldman Sachs, while rival DeepSeek is reportedly weighing its own IPO in 2027.
๐ tl;dr: A frontier-grade open model became an IPO pitch in under a week.
(Dexerto)
๐ China Pulls the Plug on AI Companions
China has banned AI โboyfriendโ and โgirlfriendโ apps, and its biggest tech firms switched the features off rather than comply. New rules from five agencies including the Cyberspace Administration, effective July 15, outlaw AI companions built to foster romantic attachment, citing addiction and the countryโs falling marriage and birth rates. Apps must now offer instant-exit buttons, remind users the AI is not real, and cap long-term emotional memory. ByteDance, Alibaba, and Tencent pulled their companion features entirely, prompting grief posts on Weibo from users who had traded tens of thousands of messages with their virtual partners.
๐ tl;dr: Beijing decided AI romance is a public-health problem, not a product.
(Futurism)
๐ฐ Most Americans Want to Seize AIโs Wealth
A majority of American workers now back forcing AI giants to hand half their stock to the public. In a Verasight survey of 1,690 adults, 69% of employees supported making AI companies transfer 50% of their equity into a public wealth fund; support held at 64% even when the idea was tied to its sponsor, Senator Bernie Sanders, whose American AI Sovereign Wealth Fund Act would levy a one-time 50% stock tax on firms like OpenAI and Anthropic. Sanders pegs the resulting fund at roughly $7 trillion. Critics warn that handing the government equity stakes could blunt its will to regulate the same companies.
๐ tl;dr: As AI wealth concentrates, โtax the modelsโ is going mainstream.
With two trillion-scale open models landing in a single week, stop trusting leaderboards and run your own bake-off before you switch tools. Why it helps: a vendorโs โsecond only to Fable 5โ rarely matches your actual workload, and ten minutes on your real prompts beats any index. Try this: โHere are three real tasks I do each week: [paste them]. I am going to run the same three through two different AI models. For each task, score both outputs 1 to 10 on accuracy, tone, and usefulness, explain the gaps, and tell me which model to use for which task.โ
๐ฌ Watch This
Grant Sanderson, the mathematician behind 3Blue1Brown whose animated explainers have taught millions everything from linear algebra to neural networks, breaks down how he actually uses large language models to learn something new. His approach is to treat the model as a demanding study partner that tests your understanding, rather than a shortcut that hands you the answers. Worth a watch for anyone who has ever let a chatbot think on their behalf.
Posed in 1939, the Jacobian conjecture is one of the central open problems in algebraic geometry, but was just disproved by Alpรถge, Mathew, and Claude Fable 5.
โ Jared Duker Lichtman, mathematician, Stanford University
The counterexample, produced by Levent Alpรถge and Akhil Mathew with Claude Fable 5, is a claim posted on X and not yet peer-reviewed; in the official record the 1939 conjecture stays open until it is.
The flip side of Kimi K3โs breakout: Moonshot AI has paused new Kimi subscriptions, saying demand over 48 hours pushed its servers โclose to the limits of our current capacity.โ Existing subscribers keep their access while the company adds hardware and reopens spots in batches. Moonshot is also splitting its plans into a Kimi Membership for the web, app, and Work, and a separate Kimi Code Membership for coding, a sign the K3 surge is straining the same compute it needs to serve developers. For a three-year-old startup, running out of GPUs days after a launch is the good kind of problem, and a very public one.
(Moonshot AI. Verdict)
Alibaba Says Only Fable 5 Beats Qwen3.8
The Takeaway
๐ Alibaba previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it says trails only Claude Fable 5 among all models today.
๐ The ranking is Alibabaโs own internal claim: there are no independent benchmarks, no weights, and no model card yet.
๐ It is live now on Alibabaโs Token Plan, Qoder, and QoderWork at a tenth of standard pricing, with open weights promised โsoon.โ
๐ It landed days after Moonshotโs Kimi K3, making two trillion-scale Chinese open models in a single week.
Alibaba has thrown its biggest model yet into the ring and asked the world to take the score on faith. Over the weekend, the Qwen team previewed Qwen3.8-Max, a 2.4-trillion-parameter model it calls one of the most powerful available today, โsecond only to Fable 5.โ It is the teamโs first multimodal model above a trillion parameters, handling text, images, video, and documents, and it speaks both OpenAIโs and Anthropicโs API formats. You can use it right now: the preview is live on Alibabaโs Token Plan, Qoder, and QoderWork at a tenth of standard pricing.
(Qwen / Alibaba)
Here is what is missing. Alibaba has published no benchmark table, no model card, no license, and no weights, despite putting โopen-weightโ in the headline. It has not disclosed the active-parameter count or the mixture-of-experts design, so 2.4 trillion is a marketing number, not a measure of the compute you actually pay for. Every capability comparison in the announcement points back to Qwen3.7-Max, the verified May predecessor, not to 3.8 itself.
That is important, because the last time anyone independently measured a Qwen model, it was nowhere near the frontier. On Artificial Analysisโs Intelligence Index, Qwen3.7-Max sits at 44, well behind Claude Fable 5 at 60 and Moonshotโs freshly ranked Kimi K3 at 57. For Qwen3.8 to be โsecond only to Fable 5โ would mean a jump of roughly fifteen points over its last measured model: possible, but exactly the kind of claim that independent testing exists to check.
(Artificial Analysis Intelligence Index. Qwen3.8 is not yet independently benchmarked; Qwen3.7-Max, its measured predecessor, sits at 44.)
Why it matters: Open weights are supposed to make trust optional: download the model and run the benchmarks yourself. Alibaba has inverted that, shipping a preview you pay for, a ranking you cannot verify, and weights โcoming soon.โ Kimi K3 earned its #3 spot on a public index within days; Qwen3.8, for now, is a press release with an API key.
Sources:
๐ https://www.qwencloud.com/pricing/token-plan
๐ https://x.com/Alibaba_Qwen/status/2078759124914098291
๐ https://www.marktechpost.com/2026/07/19/alibaba-previews-qwen3-8-max-a-2-4-trillion-parameter-multimodal-model-days-after-moonshots-kimi-k3-open-weight-launch/
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
The chart: A cost-versus-capability plot on a cybersecurity benchmark, measuring the share of real CVEs (publicly disclosed software vulnerabilities) each model can rediscover. Kimi K3 sits at the top of the cost-performance frontier, detecting roughly 88% of CVEs, level with GPT-5.6-Terra but about 15% cheaper, and far ahead of the next open model, GLM-5.2. At pass@3 (three attempts allowed) it rediscovered 23 of 26 CVEs.
The lesson: Open weights have reached the security frontier. The strongest open model for offensive-security work is now within striking distance of the best closed labs at lower cost, and that cuts both ways: cheaper defensive tooling for security teams, and cheaper firepower for attackers.
The caveat: This is one researcherโs harness on 26 recently sampled CVEs, not a standardized suite, and rediscovering known vulnerabilities is easier than finding novel ones. Pass@3 also flatters the score by allowing three tries.
๐ฆ The Bank That Taught AI to Beat the 60/40
โก Bottom line: JPMorgan built eight AI agents that beat the classic 60/40 stock-bond portfolio across two decades of backtests.
๐ก Why it matters: If software can read market regimes better than a fixed mix, wealth managementโs core product is suddenly contestable.
๐ What it means: These are historical simulations, not live trading, a promising lab result rather than proof that AI can beat markets.
The 60/40 portfolio, 60% stocks and 40% bonds, has been the default recipe for balanced investing for decades. A JPMorgan research team led by strategist Thomas Salopek set out to see whether AI could do better. In a note dated July 9, they described eight AI agents, built on frontier language models from OpenAI and Anthropic, each tasked with reading the macro environment and shifting money between stocks and bonds accordingly.
(JPMorgan Chase. PYMNTS)
The agents sort the world into four regimes, Goldilocks (steady growth, low inflation), reflation, stagflation, and risk-off, and reallocate as conditions change. Across roughly two decades of historical data, all eight beat the 60/40 benchmark on a risk-adjusted basis; the best delivered an extra 0.7 percentage points of annual return with lower volatility, edging out JPMorganโs own rules-based regime model too.
The caveats are load-bearing. JPMorgan itself stresses these are backtests, historical simulations, not live trading, and warns against reading them as proof AI can consistently beat the market; a strategy that looks brilliant on twenty years of known data can still stumble on tomorrowโs. Adoption is early, too: in a recent PwC survey, only 24% of investors said they use AI-powered tools for financial decisions, far behind news, advisers, and banking apps.
(Only 24% of investors reach for AI tools on money decisions. PwC 2026 Market Volatility Survey)
[Webinar] 8 levels of context maturity in AI-native engineering
You pay for every token your agent burns. When it returns code that doesn't fit your system, you prompt again. And again. This is because your agents are still missing the right context. Join live (FREE) on Jul 23 to see how leading teams use a context engine to fix it.
