In partnership with
In Todayโs Issue:
๐ Tencent's Hy3 open model reaches the frontier
๐งฎ Mistral ships an open model that reads math proofs
๐พ SK Hynix chases a record $29B Nasdaq listing
๐๏ธ Microsoft lets you switch off Teams Copilot
๐ A poker AI now trades billions on Wall Street
โจ And more AI goodnessโฆ
โก The Signal
Open models keep reaching the frontier even as the loudest voices insist the whole AI economy is a bubble.
Today Tencent's Hy3 and Mistral's Leanstral 1.5 both land as open, Apache-2.0 models, and Hy3 reaches the closed frontier on several agentic and science benchmarks while activating just 21B of its 295B parameters. Capability is commoditizing downward at the exact moment the capital side keeps climbing: SK Hynix is chasing a record $29B US listing to feed AI memory demand, and our release-cadence chart shows flagship models now shipping monthly instead of yearly. Ed Zitron's viral CNBC hit says none of it pencils out. The gap between "models keep getting cheaper and better" and "the economics still do not work" is the real story of the week, and it is not closing quietly.
All the best,
Kim Isenberg
(Mistral AI)
๐งฎ Mistral Ships an Open Model That Reads Math Proofs
Mistral released Leanstral 1.5, a compact open model built for formal math. The 119B-parameter mixture-of-experts design activates just 6.5B parameters per token, accepts text and images, carries a 256k-token context window, and ships under a permissive Apache 2.0 license. It is tuned for Lean 4, the proof assistant researchers use to check theorems line by line, not for chat.
๐ tl;dr: An open, math-native model aimed at formal verification, not conversation.
(SK hynix)
๐พ SK Hynix Chases a Record $29B Nasdaq Listing
SK Hynix wants to raise about $29 billion in a Nasdaq listing as soon as July 10. The world's second-largest memory maker supplies roughly 60% of the high-bandwidth memory (HBM) that feeds AI accelerators, and if the sale prices as expected it would be the largest American depositary receipt (ADR) debut on record, topping Alibaba's $21.8B in 2014. The move courts US investors who want direct exposure to the AI-memory boom.
๐ tl;dr: The picks-and-shovels of AI want to trade where the AI money already is.
(Windows Latest)
๐๏ธ Microsoft Caves, Lets You Switch Off Teams Copilot Mid-Meeting
After a backlash over always-on AI, Microsoft is adding an in-meeting toggle to Teams. Organizers and presenters will be able to turn Copilot, Facilitator, and Intelligent Recap on or off during live meetings, rather than accept them by default. The switch rolls out through July across Windows, Mac, mobile, and web, and follows complaints that Facilitator would quietly listen and post answers in the chat.
๐ tl;dr: "AI on by default" hit a wall, and user control is becoming the feature.
Before you renew another frontier-model subscription, run your real workflow through a top open model and measure the gap yourself.
Why it helps: Open models like today's Hy3 and Leanstral now match closed ones on many everyday tasks, so the only way to know whether you are overpaying is to test your own prompts, not read a leaderboard.
Try this: Take three real tasks you paid an AI to do this week. Re-run each on a leading open model through a hosted playground or API, score both 1-5 on quality, and note where the open model tied. Downgrade or cancel wherever the gap is zero.
๐ฌ Watch This
Ed Zitron takes the AI bull case apart on CNBC's Squawk on the Street, and it is the sharpest bear argument you will hear all week. His claim: OpenAI and Anthropic burn cash that climbs in step with revenue, so there is no margin story and, in his words, the field is "a $10 to $30 billion TAM industry pretending to be a $1 trillion one." Watch it to stress-test your own assumptions. Even if you think he is wrong, he names the exact things that would have to break for the boom to unwind: margins that never improve, data-center debt drying up, and the first hyperscaler to blink on capex. Nine minutes, no jargon, and a sharp counterweight to a week stacked with shiny model launches.
AI is the largest infrastructure build-out in human history.
โ Jensen Huang, CEO of Nvidia, speaking at Davos 2026
(Jensen Huang. Chesnot / Getty Images via TechCrunch)
A developer gripe is doubling as an OpenAI roadmap leak. Builders report that Codex, OpenAI's coding agent, now ships with much higher usage limits and frequent quota resets, to the point that one widely shared post argued there is "basically no reason to pay for the Claude Max 20x plan right now." The bigger tell is what came next: asked whether OpenAI's top tier is headed for Codex, Tibo Sottiaux of OpenAI replied flatly: "Ultra will be in codex." Read it as a strong signal that a GPT-5.6-class "Pro" or "Ultra" model is being lined up for the agent, though names and timing stay unconfirmed until OpenAI ships.
Tencent Ships Hy3, and the Open Frontier Just Moved
The Takeaway
๐ Hy3 is a 295B-parameter mixture-of-experts model that activates only 21B per token, so it runs far cheaper than its size suggests.
๐ It is fully open (Apache 2.0) and jumps hard over its April preview: SWE-bench Pro 46.0 to 57.9, MathArena Apex 12.6 to 38.7, BrowseComp 67.1 to 84.2.
๐ On agentic and science tests it reaches the closed frontier: it leads FrontierScience-Olympiad at 74.8 and ties GPT-5.5 and Claude Opus 4.8 on BrowseComp (84.2).
๐ It still trails the best on the hardest coding and math: SWE-bench Pro 57.9 vs Claude Opus 4.8's 69.2, MathArena Apex 38.7 vs GPT-5.5's 85.4.
Tencent has quietly released Hy3, the full version of its Hunyuan-line flagship, and it lands as one of the strongest open-weight models yet. It uses a mixture-of-experts (MoE) design, meaning only a slice of the network fires on each token: 295 billion total parameters but just 21 billion active, plus a small multi-token-prediction layer that speeds up generation. The weights are on Hugging Face under a permissive Apache 2.0 license, so anyone can run or fine-tune it.
(Tencent / Hugging Face)
The headline is how far it jumped from the April "preview." On SWE-bench Pro (real software-engineering tasks) it rose from 46.0 to 57.9; on MathArena Apex from 12.6 to 38.7; on BrowseComp (web-browsing agents) from 67.1 to 84.2. In a blind test by 270 domain experts, Hy3 scored 2.67 out of 4, edging Zhipu's GLM-5.1 (2.51), the model that set the pace for open weights this year. It also cut its internal hallucination rate from 12.5% to 5.4%.
(Tencent / Hugging Face)
What stands out is where it meets the closed frontier. Hy3 tops FrontierScience-Olympiad at 74.8, ahead of GLM-5.2 (72.5) and GPT-5.5 (73.8), and its 84.2 on BrowseComp sits level with Claude Opus 4.8 (84.3) and GPT-5.5 (84.4). The gaps are honest, though: on the hardest coding and math the closed labs still win comfortably, with Claude Opus 4.8 leading SWE-bench Pro (69.2) and Terminal Bench 2.1 (85.0), and GPT-5.5 lapping the field on MathArena Apex (85.4 to Hy3's 38.7). Hy3's pitch is not "best model," it is frontier-adjacent capability you can host yourself.
Why it matters: Every month an open, self-hostable model gets closer to the paid frontier on the tasks most teams actually run. For anyone weighing a closed-model contract, "good enough and open" is now a real option, and the pressure that puts on closed-model pricing is the story to watch.
Scale AI support on AWS, see how July 9
Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.
You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.
The chart: Frontier Labs, Release Cadence is a Superintelligence timeline of flagship model releases from OpenAI, Anthropic, Google, xAI, and Meta between 2023 and July 2026. Each dot is a headline release, and the closer the dots sit, the faster that lab is shipping.
The lesson: The dots are bunching up. OpenAI ran from GPT-5 to a GPT-5.6 preview and Anthropic stacked Opus 4.5, 4.6, 4.7, 4.8, and Fable 5, all inside roughly a year. What used to be an annual event is now closer to a monthly drumbeat, the same acceleration behind today's twin open releases.
The caveat: Cadence is not capability. A point-release is not a new model, and counting headline drops rewards frequent renaming, so the curve tracks how often labs ship, not how much better each release actually is.
๐ The Poker AI That Now Trades Billions on Wall Street
โก Bottom line: Three ex-DeepMind researchers who built a poker-beating AI raised a $500M Series A to run the same tech on stocks and crypto.
๐ก Why it matters: It is a concrete case of reinforcement learning making real money, cutting against the "AI has no business model" story.
๐ What it means: The math that mastered bluffing under uncertainty maps neatly onto markets, and quant funds are now paying for it.
In 2017, a small team at DeepMind's Edmonton lab built DeepStack, the first AI to beat professional players at no-limit Texas hold'em. The trio behind it, Martin Schmid, Rudolf Kadlec, and Matej Moravcik, left to found EquiLibre Technologies in Prague, and last week the company raised a Series A at a $500 million valuation, led by European VC Creandum in what the firm called its largest single investment ever.
(Martin Schmid, EquiLibre Technologies)
Their bet is that poker and markets are the same kind of problem: decision-making under uncertainty, scored by a brutally simple number. "The nice thing about trading and markets is that the scoring is super simple: how much money did the agent make?" says Schmid. EquiLibre's reinforcement-learning agents, models that learn by being rewarded for profit, now trade billions of dollars a day across the S&P 500 and Nasdaq in partnership with quant firm Tower Research Capital.
(Illustration: StartupFortune)
The track record is the pitch: the company claims a perfect record of zero negative months since its 2025 crypto rollout, now extended to equities. With just 25 people, EquiLibre is a reminder that the most bankable AI right now may not be a chatbot at all, but a narrow agent pointed at a domain where winning is unambiguous.
HR and IT need to work as one. Here's how
HR and IT misalignment costs you more than time. Get the guide that closes the gaps across the full employee lifecycle.
