Nvidia opens Nemotron 3.5 Super VL: a 120B multimodal agent model
Weights, data and recipes are all public; a 24GB Bolt sibling runs local. Plus: ChatGPT Pro gets parallel agents, and Postgres 19 slips to Nov 19.
Copy markdown
Nemotron 3.5 Super VL: an open 120B agent, one H200
Nvidia dropped a 120B-total / 12B-active multimodal model — text, image and video in, up to 1M context — under the permissive OpenMDW license, and released the datasets, training recipes and RL environments alongside it. The NVFP4 build runs on a single H200, GB200 or DGX Spark, and Nvidia claims up to 5x the single-GPU throughput of comparable open models on agent tasks. Pull the weights today and point your coding, computer-use or document agents at it.
Bolt 74B: a Nemotron that fits in 24GB
The same drop ships Nemotron 3.5 Bolt (74B-A9B) with native 2-bit quantization so it fits in 24GB — one RTX card or a DGX Spark — for local agents, with weights on Hugging Face now. A 4GB Bolt E4B aimed at voice and embedded agents is slated for Oct 23.
GPT-6 Pro now runs agents in parallel
OpenAI switched on parallel agents for GPT-6 Pro in ChatGPT: multiple agents take different parts of one task at once, and you can follow each one's progress and open finished work to review it. Pro also gets faster responses and better tool and plugin use. It's the ChatGPT app, not the API — no pricing or limits published yet.
Postgres 19 slips to November 19
The Postgres team pushed GA to Nov 19 and inserted an RC2 on Nov 12, calling it an 'atypical' release cycle after a late wave of back-branch fixes; secondary reports tie the churn to LLM-assisted review surfacing dozens of defect reports and several feature reverts. If you'd penciled in the late-September window for an upgrade, reset it.