OpenMDW-1.1 weights on Hugging Face, not a closed API flip. On 15 October 2026 NVIDIA posted Nemotron 3.5 Super VL and Nemotron 3.5 Bolt for agent stacks that need multimodal input on infrastructure the buyer controls.
NVIDIA opens Super VL and Bolt on the same day
NVIDIA's 15 October 2026 blog says Nemotron 3.5 Super VL is an open 120-billion-parameter model for specialized always-on agents, with 12 billion active parameters in a mixture-of-experts design. The Hugging Face BF16 card lists a Mamba-2-Transformer hybrid MoE with a C-RADIO vision encoder, context up to 1M tokens, text/image/video inputs with text output, native multi-token prediction, and OpenMDW License Agreement version 1.1. Deployment notes on that card include 4× H100 or 2× GB200 for the BF16 reference; an NVFP4 checkpoint is published for lower-precision serve.
The same day NVIDIA posted Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic work, using native 2-bit quantization to fit in 24GB and run beside other applications. The Bolt card lists 74B total and 9B active parameters, context up to 256K tokens, the same OpenMDW-1.1 license, and support paths that include llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. A smaller Bolt E4B 4GB SKU is listed for 23 October 2026. Super VL and Bolt join Lightning and Ultra in the Nemotron family framing on the blog.
vLLM's 15 October post announces day-0 support for Super VL, cites the 120B/12B MoE split, and shows example vllm serve lines for the NVFP4 repo with --reasoning-parser nemotron_v3 and MTP speculative decoding. Example hardware named on that post includes 1× GB10, B200, B300, RTX PRO 6000, H200, and 2× H100.
NVIDIA's blog states Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. It also says open datasets and NeMo Gym environments are shipping on Hugging Face. Company marketing on the same page claims up to 5× faster throughput versus other open models in its class on Artificial Analysis Intelligence Index workloads and names customers post-training the model; those throughput and customer claims are NVIDIA's.
Who the OpenMDW download binds
NVIDIA shipped the weights, license, datasets pointer, NIM packaging claim, and partner list. Operators who download Super VL or Bolt accept OpenMDW-1.1 and choose their own serve stack. Teams that follow the vLLM cookbook are bound to the pinned image and flags in that post. Buyers waiting on Bolt E4B are bound to the 23 October date on the NVIDIA blog, not a shipping confirmation in this Dispatch.
Multimodal open MoE plus a 24GB local Bolt SKU
A public Super VL multimodal open-weight SKU with day-0 vLLM recipes, plus a 24GB Bolt local SKU under the same OpenMDW-1.1 family, spanning data-center and desktop NVIDIA hardware named in the posts. It is not a closed frontier chat default. It is not a claim that every open competitor is slower on your cluster.
5× and PinchBench stay on NVIDIA's page
Date: 15 October 2026 for Super VL and Bolt 74B; Bolt E4B listed for 23 October. License: OpenMDW-1.1 on the Hugging Face cards. List API token prices are not the product here; self-host cost is your GPUs and power. Region and residency are whatever your deployment is. Independent reproduction of the 5× throughput chart, PinchBench #1, and BrowseComp #2 claims is not in the launch posts beyond Artificial Analysis / NVIDIA framing. Whether your agent harness matches company-reported SWE-bench Pro or AA index rows is unsettled until you run them.
Pin the HF repo and license before you swap the default agent
Platform owners should download the BF16 or NVFP4 Super VL id they intend to serve, read OpenMDW-1.1, and pin the vLLM image from the day-0 post before flipping any production agent default. Edge and laptop teams should treat Bolt 74B-A9B 24GB as a memory-bound local option and wait for E4B if 4GB is the constraint. Ignore "5× faster" as a purchase order until your own task mix matches the Artificial Analysis setup NVIDIA cites.