AI developer tools · What shipped

NVIDIA released several updates to its Nemotron model family and supporting tools. Nemotron 3 Nano (30B total / 3B active parameters) features a hybrid Mamba-2 + Transformer MoE architecture with 1M-token context and offers 4× faster inference than its predecessor. Nemotron 3 Super (120B total / 12B active parameters) delivers up to 5× higher throughput using a LatentMoE architecture, while Nemotron 3 Ultra (550B total / 55B active parameters) targets demanding multi-agent and reasoning workloads. For code-related benchmarking, NVIDIA released Nemotron-SFT-Competitive-Programming-v2 and Nemotron-SFT-SWE-v2 datasets alongside reinforcement learning code datasets. NVIDIA Cosmos 3 launched as an omni-model for physical AI with three variants: Super (32B class), Nano (8B class), and Edge (4B class), using a Mixture-of-Transformers architecture for unified vision reasoning and action prediction. The company also released SPEED-bench for evaluating model performance and expanded community collaborations with Llama-3.1-Nemotron variants (Ultra, Super, Nano) and Mistral-NeMo (12B with 128k context window).

Epoch and METR released MirrorCode, a benchmark for measuring AI systems' ability to complete long-horizon programming tasks. The benchmark tests how well AI can re-implement software programs based solely on CLI access without source code. Claude Opus 4.7 solved a task in 14 hours for $251 that would take humans 2-17 weeks, with 17 of 25 target programs achieving perfect scores and 21 of 25 achieving 99% or higher. Leading models from a year ago would have scored about 30%, demonstrating rapid AI model improvement over time. The benchmark consists of a scaffold and 22 target programs totaling 132 task instances across six languages. Anthropic demonstrated that Claude Opus 4.7 can autonomously complete robot tasks 20 times faster than previous human records. In August 2025, Opus 4.1 was unable to complete tasks, but by May 2026, Opus 4.7 completed all but one task in 9 minutes 35 seconds, compared to 181 minutes for humans with model assistance. Sunday released ACT-2, a robot foundation model achieving 99.1% success rate on clothing folding tasks across 9 garment types, with the startup emphasizing that scaling pretraining paired with minimal high-quality in-house data is the recipe for solving robot generalization.