Ai2

Ai2 released HiRO-ACEv2, an AI climate emulator that it says generates global 3 km data at about 1/300th the cost of the physics-based X-SHiELD model. The new HiROv2 downscaling model predicts multiple fields at once and, per Ai2, matches X-SHiELD statistics to within 3% even for one-in-a-million events.

Ai2 has open-sourced AstaBrief 8B, a model fine-tuned from Qwen3-8B that turns a research question and retrieved literature excerpts into a cited report. It is live in Asta as Fast mode, which averages 51.1 seconds per report versus 178.5 seconds for the Claude-powered Thinking mode, according to Ai2. The weights and training data are public.

Ai2 released Olmo-core 3, an open training framework for large mixture-of-experts models. The company reports about 2.7× the throughput of its earlier FSDP-based implementation and benchmarks up to 1.2 trillion total parameters on 512 NVIDIA B300 GPUs. It is the infrastructure for Ai2's next MoE-based Olmo, not a model release.

A new robotics benchmark called StationeryBench shows OpenAI's GPT-6 Astra completing 7 of 100 desk-object manipulation tasks versus zero for Ai2's MolmoAct2, with a median progress score of 46 against 12. Cornell/DeepMind researcher Yoav Artzi calls the result a 'step change in spatial reasoning.'

Ai2 has released BenchMIRT, a technique that uses multidimensional item response theory to analyze which underlying capabilities drive scores on individual benchmark questions. Trained on 100 LLMs across 16 benchmarks and 34,000+ questions, it found that benchmarks like BBQ and WMDP measure general reasoning more than safety, despite being marketed as safety evaluations.

Ai2's new TutorMoments framework replays real tutoring transcripts to test whether LLMs make the right pedagogical call at key decision points. Across seven models tested, all defaulted to over-helping unless explicitly prompted about the scaffolding-versus-rigor trade-off.

Ai2 has launched the OlmoEarth Platform, infrastructure built to run its OlmoEarth Earth observation foundation models at continent scale. A North America wildfire-risk map used nearly 20,000 CPUs and 994 GPUs simultaneously, cutting an estimated 4,737 serial-compute hours to about 30.5 hours.