Ask around about the big AI stories of 2026 and you'll get a predictable shortlist. Meta abandoning open weights, then sort of walking it back. The OpenClaw saga that broke GitHub. Jensen Huang making his first post on X, ever. All legitimate picks. All heavily covered.
My answer is none of those. The most overlooked event of the year started in late August and is running as you read this: the Marin project out of Stanford is pretraining a 535-billion-parameter model with every metric streaming to the public in real time. It's the largest fully open, live-streamed frontier training run anyone has ever attempted. Total mainstream coverage so far: a few niche blog posts and one line in a newsletter digest.
Big and invisible at the same time. That's rare enough to be worth a closer look.
So what is this thing
Marin got started at Stanford in 2024, founded by David Hall and Percy Liang. The engineering work now happens under a nonprofit called Open Athena, with compute from Google's TPU Research Cloud and backing from the Huang Foundation, the Siegel Family Foundation, and Schmidt Sciences. Since around August 24, 2026, the project has been pretraining a frontier-scale model, and doing it in a way nothing at this size has been done before.
I want to be precise about "open" here. Every experiment begins as a GitHub issue stating its hypothesis up front. Configurations land as pull requests that outsiders can review while the run is in progress. Training metrics go public the moment a run starts. Data, code, recipe, final weights: all of it stays open afterward (Aoyii). Liang announced it himself, with typical understatement: "Marin 535B-A23B started training this week! As usual, the whole process is open" (X). His backers go further and call it "the largest open and live training run in history" (Open Athena) — and for once that's not puffery.
The numbers, with a small honesty note
The model is a 535.3-billion-parameter mixture-of-experts with 22.76 billion parameters active per token. The name just rounds those two figures. Exact specs live in the tracking issue.
Token budget comes with a footnote. You'll find two official figures: 18.75 trillion tokens in Liang's announced plan (80% pretraining, 20% midtraining), but 18 trillion in the repo config and the preregistration. The homepage's own target reads "Paloma eval loss, preregistered at 18T tokens: 2.04" (marin.community). Independent write-ups have flagged the same split (LamjinLab; Aoyii). Both numbers trace back to the project itself, so treat this as a loose end rather than a scandal. It's also the kind of discrepancy that only surfaces when you do everything in public, which is kind of the point.
Hardware: eleven NVIDIA GB200 NVL72 racks, roughly 792 GPUs, rented from CoreWeave. About 2.7×10²⁴ FLOPs over a ~100-day trajectory, projected to wrap around December 1, 2026 (KAD8; 36kr). As of Liang's early-September update, the run was 13% complete (X).
Follow the money, because it leads somewhere awkward
The compute is funded by the Jen-Hsun and Lori Huang Foundation. Yes, that Huang. Liang credited them directly: "This hero run would not be possible without the generous support of the Jen-Hsun and Lori Huang Foundation, which provided the funding for the compute (Coreweave)" (X).
There's independent corroboration, sort of. A regulatory filing reported by Reuters says the foundation has purchased roughly $108.3 million of CoreWeave computing capacity for academic and nonprofit AI and scientific research (Yahoo Finance). Two caveats: the filing predates the run and never names Marin, so the link to this specific run rests on Liang's public credit. And the mechanism matters — it's a capacity purchase (unlikely a hardware gift).
There's also a conflict-of-interest thread worth pulling. NVIDIA itself invested $2 billion in CoreWeave and signed a $6.3 billion capacity guarantee, so the money behind this run sits close to NVIDIA's own commercial interests. The blunt take in the coverage, that this "isn't purely philanthropic," seems fair (AI Weekly). You can hold that thought and still note that the compute is real and the run is genuinely open. Both things are true at once.
They did the homework in public too
A 535B run sounds reckless for a nonprofit, but the groundwork was done the same way as everything else. Before the hero run, the team climbed a four-rung scaling ladder from 1.6B-A61M (48B tokens) up to 27.7B-A1.2B (926B tokens), burning about 1% of the final compute budget. On top of that sits Delphi, their scaling suite, which forecasts pretraining outcomes to within 0.2% error (RuntimeWire).
The openness has already paid off once, too. Early in the run, the community spotted a loss spike near training step 500 and traced it to router_bias in real time, straight from the public metrics (KAD8). At a closed lab, that's an internal incident report nobody outside ever hears about. Here it was a public debugging session.
The announcement also traveled. Liang's post passed 800,000 views and drew an endorsement from Andrew Ng, who called it "a precious demonstration" of openness and noted that public release of AI research "used to be the industry norm" (Aoyii; 36kr). One flag before you quote him: those lines circulate via relay pages rather than anything first-hand.
Why nobody covered it
Here's my theory, and I keep coming back to it. Every other major 2026 story is a release. A model ships, benchmarks get quoted, articles get written. That's the template.
Marin isn't a release. It moves the frontier from open product to open process, and release-shaped coverage has no slot to put it in. So it landed in niche press, got one digest mention (The Neuron), and vanished from the conversation.
The sharpest proof of the blind spot comes from inside Stanford itself. HAI published an op-ed on August 4, 2026 arguing that open-weight models aren't enough and the field needs truly open source AI, citing DeepSeek and Qwen (Stanford HAI). It did not mention Marin, which was weeks from starting the largest truly open training run ever attempted, down the hall, so to speak.
Marin is mid-run — 13% as of early September, with December 1 a projection rather than a promise. The tracking issue notes that up to ~25% token-budget degradation would be handled by shrinking the token horizon and retiming the learning-rate decay, and even friendly coverage calls the schedule "an operating plan rather than an immovable contract." Wait for the weights before declaring victory.
On the money: confirmed at the program level, inferred at the run level. The Huang Foundation's $108.3M of CoreWeave capacity is real per the filing; the tie to this run rests on Liang's statement.
Two source-quality notes. KAD8 calls Liang a "Simile AI founder," which is garbled — he's a Stanford professor and Marin's co-creator alongside David Hall, and per his ICLR bio also a co-founder of Together AI and Simile AI. And the Andrew Ng quotes should be attributed carefully, since they only circulate through relay pages.
Also, OpenEuroLLM's July 31 first-models milestone was scheduled, but nothing I found confirms a release actually happened (OpenEuroLLM). Unshipped until proven otherwise.
Why this is my pick
If I only get to name one overlooked story for 2026, it's Marin, and the reasoning is simple. OLMo 3 set the openness bar in November 2025 (Ai2). Marin is the first 2026-native attempt to carry that ethos to frontier scale, and it says so itself, crediting BLOOM, Pythia, LLM360, and OLMo as lineage. The smaller runs beneath it tell the same story: Marin 8B was the first fully open JAX model back in July 2025, a 32B followed that October, and the project landed an ICLR 2026 invited talk.
It's the only 2026 event that changes what "open" will mean in 2027, rather than just participating in the current meaning. And it's happening on roughly 792 GPUs in full public view, every metric there for anyone curious enough to look, while almost nobody looks.
That's what makes it the answer. The most overlooked story isn't the one that got buried. It's the one running in plain sight.
If you're following the run yourself, I'd genuinely like to hear what you're seeing in the curves!
Top comments (0)