Highlights from Ai2 at NVIDIA GTC 2026

Ai2 was at NVIDIA GTC 2026 last week, and through panels, livestreams, expo floor demos, and partner events, we kept coming back to the same principles: reproducibility, transparency, and practical utility.

In case you missed it, here’s a quick recap of the key moments.

Open models: beyond releasing weights

We participated in several panels on the role of open models in the AI ecosystem. Hanna Hajishirzi joined Open Models: Where We Are and Where We're Headed and Build Trust and Discovery Through Open-Source AI in Research, and Ranjay Krishna joined The State of Open-Source AI.

Throughout the week, we emphasized a view of openness that goes beyond simply releasing model weights. With Olmo, our aim has been to expose the full development process – data, code, checkpoints, and evaluations – so researchers can examine how the system was built and modify it for their own uses. A released model is just a snapshot—sharing the process behind it is what gives others something robust to build on.

Our panel discussions also underscored why this matters beyond language models. In multimodal AI, many web-scale image–text datasets don't actually describe visual content in enough detail, which makes it harder to understand why models succeed or fail. Better data, open evaluation, and visibility into the pipeline aren't secondary concerns—they're part of how shared progress happens.

Olmo Hybrid and Open Coding Agents

We also joined the NVIDIA Developer Community livestream to discuss Open Coding Agents and Olmo Hybrid.

The Open Coding Agents conversation focused on agents that developers can run locally and adapt to private or internal codebases. SERA, the family's first release, was built to make strong coding agents easier to study, customize, and deploy against private repositories.

On the Olmo Hybrid side, we walked through the research motivation behind hybrid architectures and the results so far: a fully open 7B model family pretrained on 6 trillion tokens that reaches the same MMLU accuracy as Olmo 3 using 49% fewer tokens.

A recurring theme was that access to the research pipeline changes how teams engage with models—fine-tuning, running independent evaluations, and inspecting model behavior more closely.

You can watch the livestream, try SERA, or download Olmo Hybrid on Hugging Face.

Live fine-tuning at Lambda's booth

On the expo floor, attendees could watch supervised fine-tuning of Olmo Hybrid running live at Lambda's booth. The demo offered a real-time view into both model progress and system health—observability metrics like GPU temperature, memory usage, and utilization alongside live loss curves and restart logic designed to detect faults, isolate problematic nodes, and resume training without manual intervention.

“We wanted to show people that even though [Olmo Hybrid] is a different architecture, it's also really easy to work with and build on top of,” Caia Costello, an Applied ML Researcher at Lambda, said from the booth, “and I think that's the spirit of open source.” 

For more on the infrastructure behind the collaboration, see Lambda's writeup.

Asta AutoDiscovery at the Cirrascale booth

We demoed Asta AutoDiscovery at the Cirrascale booth. AutoDiscovery explores datasets, generates hypotheses, runs experiments, and surfaces findings that might not emerge through manual analysis alone. Since launch, researchers have generated over 20,000 hypotheses spanning oncology, climate science, marine ecology, and more—and we've extended access with refreshed credit allocations. Try AutoDiscovery if you haven't yet.

We were also thrilled to be invited to co-sponsor Cirrascale’s Private AI Party to cap off our busy Wednesday – alongside Google, Cisco, and Dell Technologies – offering another setting to continue conversations with people working in AI infrastructure, deployment, and research.

Robotics and embodied AI

Our GTC panels also broadened into robotics and embodied AI, where open training data and evaluation are still comparatively limited. Simulation is becoming an important way to close that gap—generating diverse training data at scale without relying on expensive manual demonstrations.

That perspective ties directly to our recent MolmoSpaces and MolmoBot releases, which aim to make strong robotics systems not only capable but also understandable and improvable in the open.

Looking back on the week

Our time at GTC 2026 spanned panels, a developer livestream, floor interviews, live demos, and partner events. We shared work on model development, coding agents, hybrid architectures, training infrastructure, multimodal systems, and scientific AI—and we brought our transparent, truly open approach to one of the biggest stages in the AI industry today.