Highlights from Ai2 at NVIDIA GTC 2026
Allen Institute (AI2)
Ai2 spent GTC 2026 pushing open AI everywhere—panels, livestreams, live demos on the expo floor. The pitch wasn't just free weights, but the whole recipe: data, code, evals, all out in the open.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ai2 showed up to NVIDIA's GTC 2026 with a message that sounds almost old-fashioned in an industry obsessed with closed frontier models: show your work. Across panels, a developer livestream, and multiple expo booths, the Allen Institute team argued that open source AI means more than tossing model weights onto Hugging Face and calling it a day. Researchers like Hanna Hajishirzi and Ranjay Krishna made that case directly in sessions on trust, discovery, and the state of open-source AI, pointing to Olmo as the proof point — a project where the training data, code, checkpoints, and evaluations are all visible, not just the final artifact.
The technical highlight was Olmo Hybrid, a 7-billion-parameter model family trained on 6 trillion tokens using a hybrid architecture. Ai2 says it matches Olmo 3's MMLU accuracy while using 49% fewer tokens to get there, a meaningful efficiency gain if it holds up under independent scrutiny — which, notably, it can, since the whole pipeline is public. Alongside it, Ai2 introduced SERA, the first release in a new family of open coding agents built to run locally against private codebases, aimed at developers who don't want to hand their internal repos to a black-box cloud agent.
On the expo floor, Ai2 partnered with Lambda to run live supervised fine-tuning of Olmo Hybrid in public, complete with GPU temperature readouts, memory usage, loss curves, and automated fault-recovery logic that isolates bad nodes without a human stepping in. Lambda's Caia Costello put it plainly at the booth: despite the unfamiliar architecture, the model is easy to build on, and that ease is the whole point of open source. Over at Cirrascale, Ai2 demoed Asta AutoDiscovery, a system that mines datasets for hypotheses across fields like oncology and marine ecology; the tool has already generated more than 20,000 hypotheses since launch, and Ai2 is now handing out fresh credits to get more researchers using it.
The conversation also stretched into robotics, an area where open training data is still scarce compared to text and image models. Ai2's answer leans on simulation to generate diverse training scenarios without the cost of manual robot demonstrations, a strategy that connects directly to its MolmoSpaces and MolmoBot releases. The throughline for the week, across coding agents, hybrid architectures, and embodied AI, was consistent: Ai2 wants systems that are not just usable but inspectable, at a moment when most of the industry is moving the opposite direction.
My take — AI-written commentary, not fact-checked reporting
I'll say it plainly: releasing weights without the data and training recipe is marketing dressed up as openness, and Ai2 is one of the few labs actually calling that out on stage instead of just in a blog post. Olmo Hybrid's efficiency numbers matter less than the fact that anyone can go check them, which is the whole point closed labs conveniently skip. If Europe wants real AI sovereignty instead of just cheering for whichever US lab open-sources a checkpoint this quarter, this is the model — full pipeline, no asterisks — worth actually funding.
Read more about this at: Allen Institute (AI2)
