# Nemotron 3.5 Bolt: Specs, Benchmarks and Size (2026) By Ossian Kettunen · 2026-10-31 · Source: https://www.activepieces.com/blog/nemotron-35-bolt-specs-benchmarks-and-size-2026 --- A new model release matters to automation builders less for its benchmark scores than for what it does to the boring parts of a workflow: how often a step returns malformed output, how long a classification call takes, and how much it costs to run a thousand times a day. Nemotron 3.5 Bolt is worth testing on exactly those terms. Teams that wire models into ticket triage, lead routing, or document extraction, whether in custom code or in a flow builder like [Activepieces](https://www.activepieces.com), should start by swapping it into one low-risk step, comparing its outputs against the current model on a few hundred real inputs, and only then deciding whether the change deserves a wider rollout. ## Nemotron 3.5 Bolt was released by NVIDIA NVIDIA published Nemotron 3.5 Bolt on Hugging Face under the model name NVIDIA-Nemotron-3.5-Bolt-E4B-4GB, and the [model page](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) is the only source this article treats as settled. Anything the page doesn't state, such as benchmark claims or task fit, stays unverified here until you test it. ### What Nemotron 3.5 Bolt is It's a model from NVIDIA's Nemotron family, and the repository name is one of several specs I take from the page for this section. For an automation team, that means the useful questions are the ones a spec sheet answers: what input it accepts, what output it returns, and under what licence. Read the page for those, not a summary of it, including this one. ### Where to get Nemotron 3.5 Bolt weights Getting the weights is the easy part, since they live on that Hugging Face page. Reaching the model from a workflow is the harder one, and it depends on whether your automation platform has a step that can call it. ![Activepieces workflow builder with HTTP request step selected and its settings panel open showing method, URL, headers, and…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9e36e973-de4a-435f-8912-12e666216b2d/self-host-mistral-ai-enterprise-deployment-guide-3f7fe28d.webp) ### Getting an open-weight model into any automation tool Whatever platform you use, the model has to answer over HTTP before a workflow can reach it. There are three ways to get there, and they work the same way in most tools. The first is to serve the weights yourself. Ollama, llama.cpp and vLLM can each load a GGUF or a set of weights and expose an OpenAI-compatible HTTP endpoint. Any platform with an HTTP request step or a custom-model step can then call that endpoint like any other API. The cost is that you run and maintain the server. The second is to use a hosted provider that lists the model. Check the provider's catalogue for the exact model name, since a similar-sounding model is not the same one. If it is there, you get an API key and an endpoint without running hardware. The third is to check whether your platform accepts a custom base URL for its AI or OpenAI-style step. If it does, you can point it at either of the first two options and keep the platform's own model step. If it doesn't, the plain HTTP step is the fallback. Each route ends in the same place, an endpoint that accepts a request and returns a completion. Until one exists, no workflow tool can use Bolt, however good its connector list looks. ![Workflow sending a request to an endpoint and receiving a completion](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/9f05fae8-83cd-45b9-bbb6-2a6bcf7c84b5/nemotron-3-5-bolt-specs-benchmarks-and-size-2026-9502b316.webp) [Activepieces](https://www.activepieces.com) runs the model you already chose on your own provider key, so spend lands on your provider account, and MoneyGram, Moneypenny, Alan and FundingSocieties run it in production. Its "pick Bolt from the list" step only works once an endpoint like the ones above exists and your provider entry points at it. 502 integrations is what the library below lists, as connector libraries advertise big catalogues. Most of those are Gmail-style app connectors, so the count says nothing about whether an integration accepts a model endpoint. The filter sidebar's Artificial Intelligence category is the part that counts. ![Activepieces connectors library page showing 502 available integration pieces with filtering options and sample connector…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/4edd1db5-aeb0-4a8b-ad9e-9b509394e8ac/power-automate-migration-guide-fixing-connector-4cf0cd0d.webp) Check that category for the provider you need before you plan a test. ## What makes new models matter to automation teams For an automation team, a new model matters only as a candidate to swap into a step it already runs. NVIDIA's own benchmarks for Bolt E4B against Nemotron 3 Nano 4B are the evidence on offer. ### What NVIDIA's Nemotron 3.5 Bolt benchmarks show On the [NVIDIA model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB), BFCL-v3 rises from 59.07 to 67.31 percent. That's a modest gain in calling the right function, so a step that already works may not change much. A bad guess can file the wrong ticket, which is why knowing when to call a tool, ask a question or decline matters. ### How to test it on your own tasks None of these tests is your ticket queue. For classifying tickets, compare label agreement with your human-tagged history, and how often the model picks "unsure" instead of a wrong category, so you can see whether the model is too cautious or too confident. For extracting invoice fields, compare exact-match accuracy per field, especially totals, tax lines and dates, plus how it behaves on missing fields. For drafting replies, compare edit distance from what your agents actually send, and whether it invents policy or refund terms. For routing by intent, compare misroutes on ambiguous messages and whether it escalates when it should. To compare, you run the same inputs through both models, which is what the next section sets up. ## Nemotron 3.5 Bolt model size and format Bolt E4B is an **8-billion-parameter model** that NVIDIA post-trained from Google's Gemma-4-E4B-it, and the page supports little beyond that, so size and format are what you can plan around. ### Facts stated on the model page On [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16), the family runs from Super VL at 120 billion parameters, to Bolt 74B-A9B at 74 billion, to Bolt E4B at 8 billion. Bolt E4B is designed for low-memory operation, so your hardware determines which version you can use. The table below lists what the Bolt E4B page states, with anything absent marked. | Item | Fact as listed | Source | |---|---|---| | Developer | NVIDIA | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | | Base model | Gemma-4-E4B-it | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | | Parameters | 8B total, 4B effective | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | | Q4_K_M GGUF | — | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | | Release date | 10/23/2026 | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | | License | OpenMDW-1.1 (Gemma 4 base under Apache 2.0) | [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) | _Prices and plan limits checked against [huggingface.co](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB) and [github.com](https://github.com/activepieces/activepieces/pull/17037) on October 31, 2026._ Although the 4B effective count describes compute per token, you still have to hold all 8B parameters in memory, which means **the model name understates your actual RAM requirement**. Budget for that figure plus context before you point a workflow at it. ### Nemotron 3.5 Bolt license and release date In the table, the release date and license are both filled in from the page, and the license allows commercial or non-commercial use. **The license is OpenMDW-1.1, and the model is ready for commercial or non-commercial use.** The page gives a release date of 10/23/2026. For your use cases, such as field extraction or routing, Treat those as untested until you run them yourself. ## What Activepieces supports for it Activepieces counts every AI step as 1 credit when you use your own provider key, so the model charge lands on your provider account and swapping models doesn't change the platform bill. That's the extent of what we can point at. **We haven't run Nemotron 3.5 Bolt ourselves**. ### The shipped support and its date Bringing your own key, available on Plus and up from $20 a month, means the model you pick is the one you already chose, called with your own key and billed at your own provider's rate. The model cost goes to your provider bill, not ours. Reselling you a model would mean we had decided your AI strategy, and priced it, before you opened the product, so Activepieces leaves that choice with you and lets you reach the result from Claude, ChatGPT, Cursor or any MCP client. It's general model support, not a Bolt-specific integration. Whether Bolt appears in the model picker, and how it behaves on your inputs, is something to check in your own instance, not something we promise. Treat this as proof that a bring-your-own-key path exists, nothing more. It says nothing about output quality, tool-call reliability, or how Bolt handles your document formats, and we have no results of our own to fill that gap. Because the first step needs admin rights and the rest don't, the order matters. 1. An admin adds your provider key in the platform settings, on a plan that includes bring-your-own-key. 2. Paste your own API key into that provider entry. Billing and rate limits then follow your provider account, not a shared one. 3. In a flow, add an AI step and pick Bolt from the model list. If it's missing, your provider or instance may not expose it yet, so check before assuming the model's unavailable. 4. Run the step against real samples: actual tickets, invoices, or emails from your queue, including the messy ones. **Clean demo text will tell you nothing**. 5. Compare the outputs against what your current model produces on the same samples, and keep the cases where they differ. That comparison is the verdict. Once you have it, you know whether Bolt belongs in the flow. ## How to measure Bolt's actual cost and speed Downloading Bolt costs you nothing in per-token fees, but the page gives you no price comparison, so the real cost is your hardware and your time. Everything else here you have to measure. ### What the page says about cost You obtain it by downloading weights, since the model's published on Hugging Face. The license is OpenMDW-1.1 and the page says the model is ready for commercial or non-commercial use, so a commercial automation can run on it. Read it before anyone builds a flow around the model. On OpenRouter, hosted small open models are billed per token, so their price shows up on your invoice. Bolt's price shows up in GPU hours and whoever keeps the server running. ![A server rack with its night-shift keeper and empty mugs](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/68c79fa6-836f-4ad8-bd76-8c8d1e0d539c/nemotron-3-5-bolt-specs-benchmarks-and-size-2026-ff74a226.webp) ### What it doesn't say about speed That's a high-end desktop card. It tells you little about a modest server or a laptop. Nothing I have compares it to a hosted model or to the model you run now. | | How obtained | Price / license | Speed | |---|---|---|---| | Nemotron 3.5 Bolt E4B | Download from Hugging Face | Read the license on the model page | ~112 tok/s on RTX 5090; other hardware, measure yourself | | Hosted small open models on OpenRouter | API call | Per-token price on each listing | Measure yourself | | Your current model | Your existing provider | Your current bill | Measure yourself | Most of the table says "measure yourself", which is the honest state of the evidence. Pick one real step from your workflow and send the same fifty inputs to all three rows, and you'll see which version works best for your real data, the answer published benchmarks can't give you. ![Three bowls being taste-tested, a cookbook pushed aside](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/18c20751-55f6-4f2d-a10d-8cfadf2e30db/nemotron-3-5-bolt-specs-benchmarks-and-size-2026-8fbd3eef.webp) Record time per call, cost per call, and how many outputs you'd accept without editing. Whichever row wins on those three numbers is your answer. ## What to watch next for Nemotron 3.5 Bolt Recheck three kinds of source after 2026-10-31, once you've picked a winner on your own tasks. Each can change what your result means for the next workflow you build. - NVIDIA's model page edits, since that page is the only spec the vendor stands behind, and a changed line there can quietly change what you were told to expect. - Independent evaluations, starting with the MindStudio test, which show whether someone with no stake in the model gets results that match NVIDIA's claims. - OpenModelStats download history, which shows whether other teams are pulling the model or leaving it after a first look. None of these is a verdict, and none tells you how the model behaves on your tickets, your documents, or your field mappings. They tell you when to rerun your own test. Published on Hugging Face by NVIDIA, Nemotron 3.5 Bolt's a new model. For an automation team, treat the model page as the spec sheet and read it as one. Look at what it says the model's for and what it leaves unsaid, because an unlisted capability should be treated as untested in your workflow. You have to prove it yourself if a task you need isn't on the page. Structured output into a fixed schema and a tool call with set fields are two examples. On Hugging Face, the model sits with NVIDIA's own page as the source of record. Download from there and not from a mirror or a repost, so the files and the description you read match what NVIDIA published. If the page changes, you'll see it against the version you tested, and you can decide whether the change is worth another run. ## Related reading - [Self Host Nemotron 3.5 Bolt: Setup Guide (2026)](https://www.activepieces.com/blog/self-host-nemotron-35-bolt-setup-guide-2026) - [SAP S/4HANA Purchase Order Workflow for MS Teams](https://www.activepieces.com/blog/sap-s-4hana-purchase-order-workflow-for-ms-teams) - [Content Creation Workflow: How Successful Teams Do It](https://www.activepieces.com/blog/content-creation-workflow) ## References - [Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16) - [Hugging Face (NVIDIA model card)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB)