NVIDIA Nemotron, MiniMax-M2 Debut; LLMs Flunk ITBench

NVIDIA unveiled Nemotron 3 Nano Omni, a hybrid Mamba‑Transformer model that handles text, images, video, and audio in long contexts. It tops benchmarks for document, video, and speech tasks while delivering up to 9× higher throughput and lower cost than competing open‑weight models. Checkpoints and detailed reports are available on Hugging Face.

The MiniMax-M2 series introduces a Mixture-of-Experts language model with 229.9 B total parameters but activates only 9.8 B per token, dramatically cutting compute while maintaining strong performance on agentic coding, search, and reasoning tasks. The architecture combines agent-driven data pipelines, a scalable RL system called Forge, and self‑debugging capabilities.

The ARBITER study finds that test‑time sampling in LLMs produces reasoning trajectories that collapse into a few ‘basins’, causing majority‑vote mechanisms to pick the most stable but often incorrect answer. By modeling basin interactions using only the model’s own samples and hidden states, ARBITER’s additive evidence recovers up to 22% of the oracle headroom on math benchmarks.

A recent arXiv paper critiques recent claims that large language models can detect and report their own internal states. By comparing LLM behavior to human metacognition research, the authors show that prior evaluation methods conflate anomaly detection with true introspection, and that input‑only baselines perform equally well. The work suggests current evidence is insufficient to prove LLM metacognition.

IBM and Artificial Analysis introduced ITBench-AA, a benchmark that tasks LLMs with real‑world SRE incident response on Kubernetes. In the first release, leading models such as Claude Opus 4.7 and GPT‑5.5 topped at just 47% and 46% accuracy, leaving all frontier models under the 50% threshold. The results highlight the challenge of agentic enterprise IT tasks for current AI systems.

OpenAI partnered with Thrive Holdings and Crete to build a Codex‑powered tax agent that autonomously learns from production data. In its pilot, the system processed 7,000 returns, reducing accountant effort by about a third, achieving up to 97% accuracy, and improving correct field completion from 25% to 86% within six weeks.

Simon Willison notes that both Anthropic and OpenAI appear to have achieved product‑market fit, with enterprise customers now paying API‑level prices and rumors of Anthropic's first profitable quarter. Pricing shifts align enterprise costs with API usage, underscoring a mature market for their LLM services.

AI demand for high‑bandwidth memory (HBM) is forcing chip makers to allocate far more wafer capacity to GPUs, limiting supply for DDR and LPDDR used in phones and PCs. With only three major memory manufacturers, this bottleneck will drive up RAM costs and raise prices of consumer devices, especially low‑cost smartphones.

SQLite recently added an AGENTS.md file that defines how AI agents can interact with its codebase. The policy bans agent‑authored pull requests, but allows AI‑generated bug reports and proof‑of‑concept patches under strict guidelines, marking a shift toward managing AI‑assisted contributions.

Hugging Face unveiled delta weight sync in the TRL library, transmitting only changed model parameters instead of full checkpoints. This reduces per‑step data transfer for trillion‑parameter models from terabytes to tens of megabytes, enabling cheaper, disaggregated reinforcement‑learning training via a Hub bucket and vLLM.