NVIDIA's open Nemotron 3.5 Super VL runs agents 5x faster on one GPU
120B multimodal MoE, open weights and datasets; a 24GB Bolt sibling runs local. Plus Meta's 816B Muse Spark goes public and GPT-5.5 exits Codex.
Copy markdown
A 120B open MoE for multimodal agents
Nemotron 3.5 Super VL ships 120B params (12B active), reads text, images, docs and video, and comes with open weights, datasets and training recipes. NVIDIA claims up to 5x the throughput of same-class open models on a single GPU — point it at OCR, chart and table extraction, and tool-using agents. Live today on Hugging Face, OpenRouter, ModelScope and NIM.
Its 24GB sibling runs an agent on your own box
Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB, tops its class on PinchBench and ranks #2 open on BrowseComp — runnable now via Ollama, LM Studio, llama.cpp or vLLM. A 4GB Bolt E4B for voice and embedded agents lands Oct 23. Build this weekend: a private local agent with no API bill.
Update: Meta's 816B Muse Spark weights go public
The frontier model Meta gated earlier today now has downloadable open weights: 816B total, 42B active, 1M context and multimodal input, under the new Muse Model License. Self-hosting wants 700GB+ of VRAM, so most builders will reach it through Groq, Together, Fireworks, Baseten or OpenRouter — or pull it from Hugging Face, SageMaker and Google Cloud Model Garden.
GPT-5.5 is gone from Codex and ChatGPT
As of Oct 14, OpenAI retired GPT-5.5 across ChatGPT, ChatGPT Work and Codex on every plan — paid tiers now default to GPT-6 Sol, Free and Go to GPT-6 Luna. If your Codex setup pinned 5.5, it's been switched for you; the OpenAI API itself is unaffected.