Tops GPTs

Sullivan & Cromwell developed Agreement Analyzer with OpenAI to surface issues in draft agreements and return memoranda and proposed tracked changes. Partner John Savva says the tool aids deeper analysis while keeping sources and uncertainty visible.

Meta introduced Muse Spark Open, an 816 billion-parameter, natively multimodal open-weight model with a 1 million token context window. Meta says evaluations show superior performance and notes the model can run locally on Macs, though heavy RAM is required.

Granite Switch 4.1 combines Decision and speech adapters into single hosted models, the authors report. Granite-Switch-4.1-8B and -30B include Decision aLoRA, Guardian aLoRA and Granite-Speech-5.0-Turbo-CTC for transcription and decision tasks.

NVIDIA, IIT Madras and AI4Bharat release Nemotron-pretraining-Indic-v1, a corpus of about 2.8 trillion tokens across 22 Indian languages, plus an open GPU-accelerated data-generation pipeline. The pipeline uses NeMo Data Designer, Gemma-4-31B-it and document-level quality checks.

Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini say their AI can iterate chip layouts to improve efficiency, aiming to reduce the need to scale data centers amid a projected 10,000x increase in compute demand over five years. They point to AlphaChip as an early example.

Third Moment Research says it runs 10,000 agents concurrently in 36–48 hour research sessions using Codex, GPT‑6 Astra, GPT‑6 Sol and GPT‑6.1 Sol. Codex identified a backtest fix that cut runtime from ~3 hours to 3–5 minutes; Daybreak is used for automated security testing.

LFM2.5-24B-A2B is a 24B-parameter reasoning model that activates 2B parameters per token to run agents on-device. The authors report competitive results vs Qwen3.6-35B-A3B on LiveClaw (48.09 vs 38.19), long-context support to 131K tokens, and fast inference.