NVIDIA releases Nemotron 3.5 Super VL and Bolt with open weights

NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt on October 15, making two related models available for developers building agents that work with text and visual inputs. The release announcement separates the available Bolt 74B model from Bolt E4B 4GB, which remains scheduled for October 23.

The launch follows the schedule in ByteForward’s coverage of NVIDIA’s local AI software updates. Today’s change is the arrival of the model weights and deployment instructions. The earlier Nemotron 3.5 Lightning release remains a separate addition to the family.

One model family with different hardware demands

Super VL has 120 billion parameters, with 12 billion active per token. It accepts text, images and video and produces text. Its context window reaches one million tokens, shared among the input, conversation, tool results and generated output. The architecture combines Mamba, attention and expert layers with a vision encoder.

NVIDIA’s NVFP4 version reduces the memory burden through quantization. Its deployment examples include one DGX Spark, one H200, two H100 GPUs or one GB200. These examples place the larger model in workstation and server territory. A long context specification does not mean every configuration will serve that entire window with multiple users.

Bolt is compressed from Super VL to 74 billion parameters, with 9 billion active, and supports up to 256K tokens. It retains text, image and video input. The compression combines a smaller architecture with quantization. Expert weights use about two bits per parameter, while other components retain higher precision. Calling the entire model a two bit network would obscure that distinction.

What the 24GB memory figure actually includes

The most useful qualification appears in section 8.2 of NVIDIA’s Bolt technical report. Its estimated 23.83GB deployment uses a batch size of one, a context length of 131,072 tokens and an 8,192 token Mamba prefill chunk. The calculation includes weights, runtime state and a speculative decoding drafter. It explicitly excludes the vision encoder and projector.

For a developer planning an agent that reads screenshots or documents, that exclusion matters. The estimate describes a particular text serving setup. It should not be treated as proof that the complete visual workflow will fit on any GPU with 24GB of memory.

NVIDIA’s GGUF instructions for llama.cpp give fuller local examples. A configuration labeled 28GB includes the vision projector and separate LiLiCorr drafter, uses 65,536 tokens of context, and targets RTX Spark with 64GB or a discrete RTX GPU with at least 32GB. A second configuration labeled 32GB raises context to 262,144 tokens and targets RTX Spark with 128GB or a discrete RTX GPU with at least 48GB. Both examples serve one sequence at a time.

Those are NVIDIA’s example configurations. They are useful starting points for matching a workload to a machine, rather than independently verified minimum requirements. ByteForward has not run the models or tested their memory use.

The runtime still needs careful matching

Bolt’s vLLM instructions specify a particular nightly build or newer, additional b12x kernels, a custom reasoning parser and a separate drafter. They also enable remote model code. That is a more specific setup than choosing a model name in an arbitrary local chat application. Developers should follow the recipe for their checkpoint and runtime.

Super VL’s DGX Spark quick start likewise specifies particular vLLM images and NVIDIA GPU support in Docker. Its published recipes should be read as hardware specific instructions, with their context and concurrency settings kept intact during an initial evaluation.

Benchmark claims depend on the checkpoint and test

NVIDIA’s local performance comparison measures Bolt at about 84 output tokens per second using its IQ2 checkpoint and LiLiCorr drafter. The AIPerf SPEED-Bench-Coding test uses llama.cpp on Windows 11 and a Surface Laptop Ultra with 128GB RTX Spark hardware. The same table reports about 97 tokens per second for Qwen3.6 35B-A3B and 101 for Gemma 4 26B-A4B, both using UD-Q4_K_M quantization with MTP. Qwen3.8 27B, using UD-Q4_K_M and DFlash, reaches about 28. Those differences make this a comparison of complete tested setups. They do not establish a universal speed winner.

The PinchBench leaderboard also distinguishes versions. At publication, Bolt BF16 has an average score of 90.4 percent, while the compressed 24GB checkpoint averages 87.8 percent. That BF16 result cannot simply be assigned to the smaller download. PinchBench measures success on OpenClaw tasks using automated checks and model judging, which also limits how directly those scores transfer to a different agent.

For Super VL, NVIDIA’s technical launch post presents a throughput claim of up to five times faster. Its chart uses 100K input tokens and shows Super VL on a Baseten endpoint against Inkling Small. A separate Dynamo deployment test uses 16 B200 GPUs at 50 tokens per second per user to report up to eight times as many users and 85 percent lower latency before the first token. Neither figure establishes that every agent will finish its work proportionally faster.

A useful evaluation should therefore keep the exact checkpoint, serving stack, context length and agent harness together. Task completion, memory use and response speed answer different questions. A benchmark lead is most useful when it survives the configuration a team actually intends to deploy.

What developers can obtain now

ByteForward checked the public Hugging Face repositories and their file listings. Super VL’s NVFP4 file listing contains 88 weight files totaling about 75GB on disk. Bolt has downloadable quantized weights and a separate GGUF release with its visual projector. These are public, ungated repositories with model files present. Download size and operating memory are different measurements.

The main model weights are released under OpenMDW-1.1, which permits commercial use. Redistribution requires retaining the license and applicable notices. Teams still need to assess rights and obligations for their own use, including any separate components they add.

NVIDIA also has a NIM container listing for Super VL. That software is governed by NVIDIA’s separate software and product terms. Availability of freely downloadable model weights should not be read as a promise of free managed hosting or enterprise support.

The release gives builders a concrete choice to evaluate. Super VL offers the larger context and server focused configuration, while Bolt reduces the local deployment burden. For Bolt in particular, choose the complete runtime recipe before deciding whether an existing machine has enough memory.

Illustrative GeForce RTX hardware photograph, published in 2025. Photo by Gavin Phillips on Unsplash under the Unsplash License.