Open Source AI — September 5, 2026
Two stories today, and they sit on opposite ends of the spectrum.
In Plain English: Two stories today, and they sit on opposite ends of the spectrum. A research institute backed by an Abu Dhabi university released the largest fully open set of AI models ever — six sizes, from tiny models meant for watches and phones up to a 375-billion-parameter flagship, all free to download along with the training data and recipe used to build them, explicitly designed so the same family can run on a phone, a local server, or a big cluster. On the other end, OpenAI released its newest and most powerful closed model, calling it a possible step toward true general intelligence, and revealed it took the single largest training run the company has ever done — more than 100,000 GPUs running at its Texas data center. One story is about giving frontier-adjacent capability away for anyone to run themselves; the other is about how much raw computing power it now takes to stay ahead. Both bear directly on Lucien's thesis, and the Apple Angle below explains why the gap between them is exactly where the on-device opportunity lives.
Institute of Foundation Models Launches K2 Horizon, the Largest Fully Open-Source Model Fleet Ever Released
The Institute of Foundation Models (IFM), backed by Abu Dhabi's Mohamed bin Zayed University of AI, released K2 Horizon: six models ranging from 0.9 billion parameters (built for watches and edge devices) up to a 375-billion-parameter enterprise flagship, all under the permissive Apache 2.0 license and shipped with full weights, code, training data, and methodology — what IFM's Eric Xing called "open science" rather than just open weights. The lineup is explicitly built for tiered deployment: small models for phones, a 32-billion-parameter dense model designed for local hosting, and larger sparse models for on-prem servers or enterprise clusters, so developers can prototype small and scale up without changing their workflow. For Lucien's thesis, this is about as direct a hit as it gets: a frontier-adjacent, fully open model family engineered from the ground up for exactly the edge-to-on-prem deployment ladder the thesis says is coming.
Institute of Foundation Models · Open-Weight Models · Sep 3, 2026
OpenAI Ships GPT-6 Astra, Its Largest Training Run Ever, and Calls It a Possible Step Toward AGI
OpenAI released GPT-6 Astra, a model built to act directly inside software applications rather than just suggest what to do — formatting legal documents, designing circuit boards, filling out tax forms — and president Greg Brockman told reporters "welcome to the AGI era," calling it a "generational leap." The company confirmed Astra came from its largest training run to date, using more than 100,000 GPUs at OpenAI's Stargate site in Texas; no pricing was disclosed, and safety evaluators flagged reduced monitorability during testing. For Lucien's thesis, the number that matters isn't the AGI talk, it's the 100,000 GPUs: that is the scale of capital and power a single closed frontier training run now consumes, with no visibility yet into whether the resulting API revenue will cover it.
Axios · Infrastructure & Economics · Sep 3, 2026
The Apple Angle (Standing Perspective)
Today's two stories almost illustrate the standing argument by themselves. On one side, OpenAI needed more than 100,000 GPUs in a purpose-built Texas facility just to train GPT-6 Astra — compute so specialized and expensive that only a handful of companies on earth can field it. On the other side, IFM's K2 Horizon shipped a 32-billion-parameter model explicitly built for local hosting, part of a family that scales all the way up to 375 billion parameters, free for anyone to download today. The gap between those two stories is exactly the terrain Apple's Mac Studio occupies: a single machine, configured with up to 512GB of unified memory shared between CPU and GPU, costs around $9,500 and can hold a trillion-parameter open-weight model entirely in memory — several times larger than even K2 Horizon's flagship. Reaching that same memory footprint with Nvidia's RTX Pro 6000 workstation cards takes five or six of them, roughly $60,000–$75,000 combined, while pulling on the order of ten times the power.
Training a frontier model will likely always require data-center-scale compute like OpenAI's Stargate cluster. But running one — once it's open, downloadable, and improving every few weeks the way K2 Horizon and its predecessors have — is a different problem, and Apple's unified-memory design already solves a large chunk of it almost by default, with little public attention. Apple's MLX framework and Thunderbolt 5's RDMA networking let multiple Mac Studios pool memory into a cluster for even larger models, and Nvidia's own DGX Spark — a small unified-memory box in its own right — is a tacit admission that this is the right shape for local inference, notably after Nvidia pulled NVLink pooling from its consumer and workstation cards. The fair caveat holds as always: a Mac Studio serves one user running one model, while the GPU clusters training and serving models like GPT-6 Astra are built for many thousands of concurrent users — Apple looks positioned to dominate a specific, high-value slice of AI compute, not to replace the data center outright.
Limited Edition Jonathan (Substack) · Perspective
Why This Matters
Today's two items sit on opposite sides of Lucien's thesis and reinforce both halves of it at once. K2 Horizon is a concrete, present-tense example of open-weight capability closing the gap with commercial frontier models while being engineered specifically for edge-to-on-prem deployment — not a future possibility, but a shipped model family with a 32B variant built for exactly that purpose. GPT-6 Astra, meanwhile, is a reminder of just how much capital and power the closed, rented side of the industry now burns to stay ahead, with a 100,000-GPU training run and still no public accounting of whether it pays for itself. Neither story proves the on-prem transition has arrived, but together they sharpen the fork in the road: one path keeps demanding Stargate-scale infrastructure, and the other keeps shipping models sized for a desk, which is exactly the opportunity the standing Apple Angle argues Apple is quietly positioned to capture.
The Daily: Open Source AI — a recurring research brief for Lucien Engelen
Book Lucien for your next event
Keynotes, masterclasses, panels and board-room sessions.