NVIDIA Releases Nemotron 3.5 Super VL and Bolt for Multimodal and Local Agents

Two Nemotron 3.5 models are available now; the 4GB model is not

NVIDIA announced Nemotron 3.5 Super VL and two Bolt variants on October 15, 2026. The release says Super VL and Bolt 74B-A9B 24GB are available now, while Bolt E4B 4GB is coming October 23. The first two have public, ungated Hugging Face repositories; the future date makes E4B a scheduled release, not a downloadable model that builders can validate today.

The identifiers matter because “Nemotron 3.5” is a family, not one interchangeable endpoint. The reviewed artifacts are nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 and nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB. Record the full repository and revision in an evaluation rather than relying on the family name.

Super VL targets high-capacity multimodal agent workflows

Super VL has 120 billion total parameters and activates 12 billion per token. NVIDIA describes a hybrid of Mamba-2, Transformer attention and mixture-of-experts layers, plus a C-RADIO vision encoder. The NVFP4 artifact accepts text, images and video and produces text; the model card lists reasoning, coding, tool use, document understanding and video understanding as intended capabilities.

Its context ceiling is one million tokens, shared by the prompt, visual inputs, conversation history, tool results and generated response. That number is capacity, not evidence that a million-token workload is accurate, fast or economical. NVIDIA’s own recipes vary batch and concurrency settings, and its DGX Spark quick start serves one concurrent sequence.

Bolt trades context and precision for a 24GB packed model

Bolt 74B-A9B has 74 billion total parameters and 9 billion active parameters. NVIDIA uses IQ2 two-bit quantization with quantization-aware training and lists a 24GB packed model, text-image-video input, text output and up to 256K context. The model card names Ampere A100, Hopper H100-80GB and Blackwell as vLLM deployment targets; NVIDIA also links local routes through llama.cpp, Ollama, LM Studio and Unsloth.

A 24GB packed artifact is not the same as a 24GB end-to-end memory requirement. Runtime buffers, multimodal encoders, context state, speculative decoding, the operating system, display use and agent tools all need capacity. “Run alongside other applications” remains a vendor claim until measured on the exact workstation or laptop configuration.

The license permits use but still needs legal review

Both reviewed cards label the license OpenMDW 1.1 and say the models are ready for commercial or non-commercial use. The repositories were public and ungated when checked, so obtaining the files did not require a manual access request. Public download access does not mean public-domain software, and “open model” should not be shortened to “no conditions.”

Review the actual OpenMDW agreement, downstream notices, application obligations and any separate terms for NIM, hosted endpoints or data before deployment. NVIDIA did not publish a universal per-token price for downloading and self-hosting these weights on the reviewed pages; infrastructure, engineering and support costs remain with the operator.

The headline speed result is provider- and workload-specific

NVIDIA calls Super VL the fastest open agentic model in its class and says it reaches up to five times the throughput of the next-best comparison. Its technical post shows more than 600 output tokens per second on a Baseten endpoint for the Artificial Analysis Intelligence Index with a 100K-token input and compares the point with Inkling Small. That is useful configuration evidence, not a speed promise for DGX Spark, H100, another provider, another context length or an interactive agent.

Throughput and task completion are separate measures. A high batched token rate can coexist with slow time to first token, tool latency, retries or incorrect actions. Reproduce cold and warm performance with the intended concurrency, input mix, tools and acceptance criteria before sizing a service.

The benchmark table is extensive but vendor-reported

The Super VL card reports results across reasoning, coding, instruction following, tool use, long-context, multimodal and agent benchmarks. Examples include 71.07 on SWE-bench Verified, 62.78 on Terminal-Bench 2.1, 77.5 on BrowseComp and 52.3 on OSWorld-Verified for the NVFP4 checkpoint. NVIDIA says the BF16-versus-NVFP4 comparison held harness settings constant and links NeMo Gym recipes for many evaluations.

AccessAllGPT did not reproduce those scores, and NVIDIA says several benchmarks still use official implementations or internal scaffolding it plans to open later. The results are vendor-run evidence, not independent proof of production reliability. They also do not measure an organization’s permissions, data quality, tool contracts, policy compliance or cost per accepted task.

Deployment support is concrete but tied to NVIDIA systems

For Super VL, NVIDIA lists vLLM, SGLang and TensorRT-LLM paths on Blackwell and Hopper. The card provides configurations for one DGX Spark, one H200, two H100s or one GB200, though modality and concurrency vary by recipe. It also links NIM packaging and Dynamo recipes. These are unusually concrete starting points, but they remain NVIDIA-optimized configurations rather than portable guarantees.

Pin the container image, model revision, remote code and runtime version before testing. The quick starts use trust_remote_code, nightly or release-candidate components in places, parser selections and hardware-specific cache settings. Review that executable and dependency surface under the same software-supply-chain rules as application code.

What AI builders should do next

Trial Bolt when a local 24GB-class agent is the real constraint and 256K context is enough. Trial Super VL when multimodal quality, longer context or higher-capacity reasoning justifies the larger NVIDIA deployment. Compare both against the incumbent model on accepted outcomes, not only tokens per second. Wait for Bolt E4B until its actual weights, model card, license, checksums and runtime guidance are available.

For either released model, freeze the repository revision and runtime, run representative text and visual tasks, validate tool-call schemas and permissions, test long-context degradation, record time to first token and completed-task latency, and measure peak memory, power and failure recovery. Reject the model for a workload if its quality, hardware, license or operational envelope misses the deployment contract.

Disclosures

AccessAllGPT received no model files, hardware, hosted access, API credits, benchmark data, private briefing, review, payment or compensation for this article. NVIDIA, Hugging Face, Baseten and the named runtime or hardware vendors did not sponsor, review or endorse it. AccessAllGPT did not test or rank the models. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with NVIDIA or the named organizations. Publication-wide relationships are listed on the disclosures page.