NVIDIA wordmark in black letters on a white background, for the Nemotron 3.5 models guide

Nemotron 3.5: Power-User Guide to NVIDIA’s Super VL, Bolt and Transcribe Models

by

Last updated: October 31, 2026

Nemotron 3.5 is NVIDIA’s open model family, and October 2026 brought four new members. Super VL is a 120B multimodal model for agents. Bolt 74B-A9B compresses it to fit 24GB of memory. Bolt E4B targets small local assistants, and Transcribe handles speech recognition in 40 languages.

This guide draws only on NVIDIA’s own material: the NVIDIA blog announcement, the Bolt 24GB research page, the Super VL model card, the Bolt E4B model card and the Transcribe launch post on Hugging Face, all checked on October 31, 2026.

Contents

What is new in Nemotron 3.5

The NVIDIA blog dated October 15 frames the family by job. Lightning, released in August, is the high-throughput model. Ultra is reserved for the hardest reasoning. The October releases fill in the middle and the edge:

  • October 15: Super VL 120B-A12B and Bolt 74B-A9B became available.
  • October 23: Bolt E4B 4GB was released on Hugging Face and build.nvidia.com.
  • October 28: NVIDIA introduced Nemotron 3.5 Transcribe, an open-weight speech recognition model, with the Granary v2 dataset.

NVIDIA says Super VL and Bolt are available on Hugging Face, OpenRouter, ModelScope and as NVIDIA NIM microservices.

Nemotron 3.5 model table

Model Parameters Inputs and outputs Best for Released
Super VL 120B-A12B 120B total, 12B active Text, images and video in; text out Agents that need vision; 4× H100 or 2× GB200 Oct 15
Bolt 74B-A9B 24GB 74B total, 9B active Compressed from Super VL Local agents in 24GB; llama.cpp, Ollama, LM Studio Oct 15
Bolt E4B 4GB 8B total, 4B effective Text in and out, 64K context Voice and device assistants; llama.cpp Oct 23
Transcribe About 4B active of 30B Audio in; text out Recorded audio in 40 languages; vLLM or SGLang-Omni Oct 28

Super VL: the multimodal flagship

The Super VL model card lists text, image and video input, with MP4 clips of up to two minutes, and text output. NVIDIA’s blog says the model was trained on about 26 trillion tokens and post-trained on about 3.8 million samples. It also says Super VL is five times faster than the next best model in its class, based on Artificial Analysis throughput figures.

Reasoning has three settings. Use enable_thinking set to True or False, or add medium_effort for a middle level. NVIDIA recommends temperature 1.0 and top_p 0.95 for text, and temperature 0.6 for vision tasks.

Selected scores from the card: GPQA-Diamond 89.52, SWE-bench Verified 71.10, LiveCodeBench v6 88.22, Terminal-Bench 2.1 60.67 and OSWorld-Verified 51.14. These are NVIDIA’s own measurements.

Bolt 74B and Bolt E4B for local hardware

According to the Bolt research page, Bolt 74B-A9B starts from Super VL and is compressed with a method NVIDIA calls Star Elastic. It routes to 9 experts instead of 22, quantizes the experts to about 2 bits (IQ2_XXS) with quantization-aware distillation, and stores the Mamba layers in NVFP4. The goal was at least 80 tokens per second on a 24GB machine.

A 1.13B-parameter LiLiCorr drafter speeds up decoding. NVIDIA measured 84 tokens per second with it on an RTX Spark laptop using llama.cpp. It also reports first place on PinchBench and second on WebArena and BrowseComp among models of similar footprint. Checkpoints come as a 24GB GGUF, a drafter GGUF and BF16.

Bolt E4B is a different model. Its model card says NVIDIA post-trained Google’s Gemma 4 E4B to improve multi-turn tool use, instruction following and grounding. It has 8B total parameters (4B effective), a 64K context window and a Q4_K_M GGUF file. A conversational_mode switch makes replies short and lightly formatted for speech.

Nemotron 3.5 Transcribe

The Transcribe launch post describes a promptable, open-weight speech model. It transcribes 40 languages, translates 31 to and from English, and attributes speakers in 13. It accepts recordings of up to one hour with up to eight speakers and returns word-level timestamps.

NVIDIA reports first place among open-weight models on the Hugging Face Open ASR Leaderboard for English, at 5.33% average word error rate. The model uses about 4B active of 30B total parameters. You can serve it with vLLM or SGLang-Omni, or use hosted versions from Baseten, DeepInfra and Simplismart.

How to run Nemotron 3.5

These steps follow the commands on the model cards. Download files with the Hugging Face CLI covered in our hf CLI new features guide.

  1. Pick a model for your hardware: Super VL for a multi-GPU server, Bolt 74B for a 24GB machine, or Bolt E4B for small devices.
  2. For Super VL, start the vllm/vllm-openai:v0.32.0 container and run vllm serve with --tensor-parallel-size 4, --reasoning-parser nemotron_v3, --tool-call-parser qwen3_coder and MTP speculative decoding.
  3. Confirm the server is up with curl --fail http://localhost:8000/v1/models.
  4. For Bolt E4B, run hf download nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf, then start the llama.cpp server image with -c 65536 -ngl 999 --jinja. Add the DFlash drafter with --spec-type draft-dflash for faster output.
  5. Point any OpenAI-compatible client at http://localhost:8000/v1.
  6. Pass chat_template_kwargs such as {"enable_thinking": false} to control reasoning per request.
  7. Use the recommended sampling settings: temperature 1.0 and top_p 0.95.

How it compares

Against the earlier Nemotron 3 Nano-4B, Bolt E4B scores higher on most of NVIDIA’s tests at BF16: 55.10 versus 29.20 on Tau2-Bench and 63.79 versus 60.10 on IFBench. It also beats Gemma 4 E4B-it on Tau2-Bench (38.50). Qwen3.5 4B still leads that test at 75.06.

On speed, the Bolt research page shows Qwen3.6 35B-A3B at 96.63 tokens per second and Gemma 4 26B-A4B at 100.56, both faster than Bolt’s 83.99 on the same laptop. NVIDIA’s case for Bolt is accuracy within 24GB. To run it locally, see our Ollama new features guide and LM Studio Bionic guide. For other open models, read our Mistral Large 4 guide and Liquid AI LFM3 guide. For speech, compare Transcribe with our Voxtral 3 guide.

Limitations

  • Super VL needs data-center GPUs; NVIDIA lists four H100 or two GB200 for deployment.
  • The Super VL and Bolt E4B cards list Linux as the supported operating system.
  • Bolt 74B relies on 2-bit experts, so test it on your own tasks before replacing a full-precision model.
  • Bolt E4B accepts text only, and its llama.cpp recipe is still marked pending.
  • Transcribe is built for recorded audio, not live streaming.

Frequently asked questions

What is Nemotron 3.5?

Nemotron 3.5 is NVIDIA’s family of open models. The October 2026 releases are Super VL, a 120B multimodal model; Bolt 74B-A9B, a compressed version that fits in 24GB; Bolt E4B, a small model for local assistants; and Transcribe, a speech recognition model. Lightning arrived in August.

What is the difference between Nemotron 3.5 Super VL and Bolt?

Super VL has 120B total and 12B active parameters, accepts text, images and video, and NVIDIA suggests four H100 or two GB200 GPUs. Bolt 74B-A9B is compressed from Super VL to 74B total and 9B active parameters with 2-bit experts, so it fits in 24GB of memory.

Can I run Nemotron 3.5 with Ollama or LM Studio?

NVIDIA lists llama.cpp, Ollama, LM Studio, Unsloth and vLLM support for Bolt 74B. The Super VL model card lists vLLM and SGLang on Linux, and the Bolt E4B card gives a llama.cpp server recipe with an optional DFlash drafter.

How do I turn reasoning on or off in Nemotron 3.5?

Pass enable_thinking as True or False in the chat template arguments. Super VL also supports a medium effort mode by adding medium_effort set to True. Bolt E4B adds a conversational_mode switch for short spoken replies and a reasoning budget in llama-server.

What languages does Nemotron 3.5 Transcribe support?

NVIDIA says Nemotron 3.5 Transcribe transcribes 40 languages, translates 31 languages to and from English, and attributes speakers in 13 languages. It handles recordings up to one hour with up to eight speakers, but it is built for recorded audio, not live streaming.

What license do Nemotron 3.5 models use?

The Super VL and Bolt E4B model cards list the OpenMDW License Agreement, version 1.1, and say the models are ready for commercial use. Bolt E4B is post-trained from Gemma 4 E4B, so the Apache 2.0 license also applies to the underlying Gemma model.

Sources

Image: NVIDIA wordmark, from Wikimedia Commons (public domain, simple text logo), placed unchanged on a white background. NVIDIA and Nemotron are trademarks of NVIDIA Corporation. FreeAnswers is not affiliated with or endorsed by NVIDIA.



Related answers

Leave a Comment