NVIDIA Releases Nemotron 3.5 Super VL And Two Bolt Models
NVIDIA published an announcement on October 15, 2026 expanding its open Nemotron model family. Everything below comes from that page.
Transcript
NVIDIA released new open models for agents, it says. The announcement names Nemotron three point five Super V L.
The company reports the model serves always on agents. It writes that parameters total one hundred twenty billion.
Per that page, Bolt models run locally. The blog notes those workloads never reach the cloud.
It claims task completion runs five times faster. The post explains a leaderboard supplied that ranking.
The page adds that the Super model reads documents. Pretraining used twenty six trillion tokens, NVIDIA's figures show.
NVIDIA published this itself about its own models and hardware. The speed and ranking numbers are unverified; another model is unreleased.
Inboxsmith builds AI receptionists for small businesses. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Line-by-line backing for every narrated sentence, in order, drawn only from the source post. Hook sentence one, NVIDIA released new open models for agents, it says: backed by the body line NVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model, and by the line Joining it are two new edge models, plus the standfirst description of open models for always-on agents. Hook sentence two, the announcement names Nemotron three point five Super V L: backed by the post title and by the same release line naming Nemotron 3.5 Super VL. Scene one sentence one, the model serves always on agents: backed by the line that the model delivers the fastest task completion and highest accuracy for specialized always-on agents. Scene one sentence two, parameters total one hundred twenty billion: backed by the same line, an open 120-billion-parameter model; the glued string 120-billion-parameter is quoted on screen only and the figure is voiced as words. Scene two sentence one, Bolt models run locally: backed by the line that Bolt 74B-A9B 24GB is designed to run everyday agentic tasks locally on NVIDIA RTX/DGX Spark, and by the line that Bolt E4B 4GB is built for physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark; the model-size tokens and the hardware names are on screen only. Scene two sentence two, those workloads never reach the cloud: backed by the line that users can delegate their projects to popular local agent harnesses, from researching the web to writing to debugging code, all without sending those workloads to the cloud. Scene three sentence one, task completion runs five times faster: backed by the line As the fastest open agentic model in its class, 5x faster than the next best, it helps enterprise agents complete high-volume tasks faster. The word agentic from that sentence is never voiced; narration says task completion instead. Scene three sentence two, a leaderboard supplied that ranking: backed by the chart caption that attributes the speed comparison to the Artificial Analysis Intelligence Index leaderboard. Scene four sentence one, the Super model reads documents: backed by the vision lines that strengths are targeted at optical character recognition, tables, charts and structured information in visual documents, that it processes a web page, extracts text, reads graphs and charts, and that the same applies to PDFs and other documents. Narration names a single capability rather than a list, so the three-item list stays on screen. Scene four sentence two, pretraining used twenty six trillion tokens: backed by the line that Nemotron 3.5 Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining. The companion figures, roughly 3.8 million post-training samples and the one-fifth parameter count comparison, are on screen only so that only one number is voiced here. Scene five sentence one, NVIDIA published this itself about its own models and hardware: backed by the byline Kari Briski on NVIDIA's own corporate blog, and by the fact that every product named on the page, the models, NVIDIA NIM, NeMo Gym, NVIDIA Dynamo, DGX Spark, DGX Station, RTX PRO, GeForce RTX and Jetson, is NVIDIA's own. Scene five sentence two, the speed and ranking numbers are unverified and another model is unreleased: backed by the absence of any cited reviewer, auditor, regulator or second outlet anywhere on the page for the 5x claim, the one-fifth parameter claim, the PinchBench #1 and BrowseComp #2 rankings or the up to 8x more users and 85% lower latency Dynamo figures, and by the availability line Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming Oct. 23, which means the third model cannot be obtained yet. Claims deliberately left unspoken for lack of support, each named on a NOT PUBLISHED line on screen: no price or licence cost for any of the three models, no licence name for the open weights, no context window length, no benchmark scores or absolute numbers, no named competitor behind the next best, no accuracy or error rate for the vision and document reading, no unit of measure detail behind per GB, no ship or availability detail for Bolt E4B beyond the stated Oct. 23 date, and no outside review of any adopter deployment. Nothing from the timeline probes is used, because those videos are unrelated. Voice safety audit. Exactly three numbers are spoken. One hundred twenty billion is spoken in the second sentence of scene one, five times is spoken in the first sentence of scene three, and twenty six trillion is spoken in the second sentence of scene four. Counting the twelve narrated sentences in order, two in the hook and two in each of the five scenes and two in the endcard, those numbers fall in sentences four, seven and ten, so no two numbers sit in the same sentence, no two sit in adjacent sentences, and at least two number-free sentences separate every pair. The tokens three point five inside the mandated model-name rendering in sentence two are part of a product name rather than a quantity, and sentence two is separated from the first spoken number in sentence four by a number-free sentence. All spoken figures are written as words, with no numerals anywhere in narration. The word agentic appears nowhere in the hook, in any scene narration or in the endcard narration; it survives only inside quoted on-screen text. Every banned glued or initialism token is on screen only and appears in no narration field: 74B-A9B, E4B, 24GB, 4GB, 2-bit, RTX, DGX, DGX Spark, GeForce RTX, Jetson, NIM, NeMo Gym, vLLM, llama.cpp, APEX-SWE, M&A, PinchBench, BrowseComp, R1, AI21 and Dynamo. No dotted domain brand is voiced, so llama.cpp and inboxsmith.com are on screen only. The model names are voiced sparingly: Nemotron three point five Super V L once, with V and L as separate single-letter tokens, the Super model once, and Bolt once. No narration sentence contains a list of three or more items; the longest spoken pairing is speed and ranking numbers, and every company list, hardware list, framework list and capability list is confined to on-screen text. No em-dash, no en-dash and no double hyphen appears anywhere in this file, including the on-screen blocks and this sources block; the only hyphens present are single hyphens inside quoted product and specification strings on screen. No idiom, metaphor or figure of speech appears in narration or on-screen text, and the source's stock phrasings from data center to desktop, push accuracy higher, control runs in two directions and context carries through are reused nowhere; the companion NVIDIA post is never labelled a deep dive. The word breaking and every countdown phrase, including the word expires, are absent from all copy fields. Distinct attribution phrases, one per factual sentence in order: it says, The announcement names, The company reports, It writes, Per that page, The blog notes, It claims, The post explains, The page adds, NVIDIA's figures show, and the closing disclosure sentences that name NVIDIA as the publisher of the page and state that its numbers are unverified.(Provenance note, not a real video title. The single source for this script is NVIDIA's own first-party blog post titled Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models, published on NVIDIA's corporate blog at blogs.nvidia.com under the byline Kari Briski, dated October 15, 2026 and captured on October 15, 2026, which is the same day this Short was produced and the same day the post is dated. The subjects are NVIDIA's own open models, its own software stack and its own hardware, so the page is first-party promotional material and not independent reporting. Unlike a vendor case study, it IS a genuine product launch: it carries launch verbs, it names three models, Nemotron 3.5 Super VL, Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB, and it states availability now for Super VL and Bolt 74B-A9B and Oct. 23 for Bolt E4B. Corroboration is NONE. The research step searched YouTube for supporting material and retrieved zero of five candidate transcripts, because the installed yt-dlp build on this host cannot fetch captions. The dated timeline probes that did return were unrelated videos about Ollama fine-tuning, Qwen3 and a Claude Code API error, so no fact in this script rests on them and not one of them is cited anywhere in this file. The regional screen is clean: the page names no regional newsroom category and no non-US customer, and it carries no regional availability section at all. No outside party has verified the 5x speed claim, the claim of accuracy comparable to NVIDIA's largest reasoning model at roughly one-fifth the parameter count, the PinchBench #1 or BrowseComp #2 rankings, or the figures of up to 8x more users and 85% lower latency on NVIDIA Dynamo. The long adopter and integration lists are supplied by NVIDIA, with no independent review of any named deployment. Because of that, every factual narration sentence names its attributor out loud with a different attribution phrase, and the final scene states plainly that NVIDIA published this itself about its own models and hardware, that the speed and ranking numbers are the company's own and unverified, and that one of the three announced models is not released yet.)