Technical Content Writer - Contract

Nvidia's Nemotron 3.5 Super VL. After extra training on medical insurance claims, Nvidia's newest open model jumped from about 28% to 72% accuracy on that task, according to the company's own chart. The model behind that result is Nemotron 3.5 Super VL, which Nvidia released on Thursday alongside two smaller Bolt models for local AI agents. Key takeaways. * Nemotron 3.5 Super VL is an open 120-billion-parameter mixture-of-experts model with 12 billion active parameters that reads documents, charts, images and video. * Nvidia releases the weights, datasets and training recipes, under a license that allows commercial use, fine-tuning and redistribution. * Customized versions beat Claude Opus 5 and GPT-6.1 Sol on five narrow tasks in Nvidia's charts, but no one has checked those results independently. Text and pictures in one model. Super VL pairs a hybrid Mamba-2 and attention backbone with a RADIO vision encoder. The mixture-of-experts design activates about 12 billion of its 120 billion parameters for each token. Nvidia distilled the model from Nemotron 3 Ultra, its largest reasoning model. It claims Super VL matches Ultra's accuracy on multimodal agent tasks at roughly one-fifth the size. The model targets what business agents actually see: scanned text, tables, charts, PDFs and web pages. Handling text and images in one model means an agent does not have to pass context between separate systems. Built for speed. Nvidia calls Super VL the fastest open agentic model in its class. Nvidia says independent benchmarker Artificial Analysis measured about 625 output tokens per second on four Blackwell B200 GPUs with 100,000 input tokens. That is about five times the next model in Nvidia's comparison. Raw intelligence scores tell a more modest story. Super VL lands near 26 on the Artificial Analysis Intelligence Index, below open rivals such as GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40. Nvidia's headline claim rests on five specialized tests: document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims. Post-training lifted the base model by 9 to 44 percentage points across them. In each test, the customized version scored above a named closed model. On medical claims, it reached about 72% against roughly 38% for Claude Opus 5. On security attack resilience, it rose from about 35% to 76%, against roughly 66% for GPT-6.1 Sol. These are different custom versions on different tasks, not proof that the base model beats closed systems in general. Nvidia has not published the task definitions, sample sizes or test protocol. Partners did the customizing. Scale AI tuned the model for Skydance's vendor-credit finance agent and EY's merger due-diligence agent. AfterQuery built a credit-agreement version for Hebbia, AI21 and Alice trained it for cybersecurity, and Mercor trained it for coding agents. Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys are integrating it. Bolt models for local machines. Nemotron 3.5 Bolt 74B-A9B has 74 billion parameters and uses 9 billion at a time. Native 2-bit compression squeezes it into 24GB of memory so it can run on RTX Spark and DGX Spark systems beside other apps. Nvidia ranks it first among open models in its class on PinchBench and second on BrowseComp, without giving scores. Bolt 74B-A9B is available now with llama.cpp, Ollama, LM Studio, Unsloth and vLLM support. A 4GB Bolt E4B model for voice assistants and embedded agents follows on 23 October. What "open" Covers here. Super VL trained on about 26 trillion tokens, then on roughly 3.8 million post-training samples with reinforcement learning across dozens of environments. Nvidia publishes the datasets on Hugging Face, plus the training recipes. The models use Nvidia's OpenMDW license. Super VL is available on Hugging Face, OpenRouter, ModelScope and as an Nvidia NIM microservice. Developers get a free, rate-limited API, while the production NVIDIA AI Enterprise license costs $4,500 per GPU per year. Open models, Nvidia hardware. Nvidia's open-model push doubles as a hardware strategy. Super VL is optimized for Blackwell and Hopper systems, and an independent review of the launch notes that enterprises still need to test it on their own documents before committing. The timing helps: cloud giants now design their own chips, as Microsoft did with its plan to deploy a million Maia 300 chips, so strong open models keep developers close to Nvidia systems. Distribution is the other half of the plan. Nvidia has agreed to buy Hugging Face for nearly $13 billion, a deal still awaiting approval and expected to close in the first half of 2027. Nvidia has pledged to keep the platform open to other model makers and chip vendors. Meanwhile, closed providers compete on price, with Google offering Gemini 4 Argon at $2 per million tokens.

AfterQuery becomes Y Combinator's fastest unicorn ever at $3.2 billion. AfterQuery, an 18-month-old AI training-data startup founded by three twenty-somethings, has reportedly hit a $3.2 billion valuation, ten times its April price tag and the fastest run from launch to unicorn in Y Combinator's history. The company pays doctors, lawyers, and other specialists to train AI models on expert reasoning rather than simple annotation. AfterQuery has reached a reported $3.2 billion valuation just months after its Series A, and the pace tells you how badly AI labs now want expert human reasoning data. AfterQuery is only about 18 months old. But it's already being priced like a company that sits near the center of the AI buildout, and Forbes first reported on September 1 that the San Francisco startup had reached that $3.2 billion valuation, citing two people with direct knowledge of the matter. TechCrunch later noted the same figure and said Y Combinator partner Gustaf Alströmer called it the fastest move from launch to unicorn status in YC history. That's fast. Five months ago, AfterQuery announced a $30 million Series A at a $300 million valuation. Now the price is more than ten times higher. That's before most startups from the same batch have had time to prove whether their first customers will renew. The founders. AfterQuery came out of Y Combinator's Winter 2025 batch. It lists Spencer Mateega, Carlos Georgescu, and Danny Tang as founders on its own site. YC's company profile names Mateega as founder and CEO and Georgescu as founder and CTO. Forbes described the company as founded in February 2025 by Mateega and Georgescu, after an earlier plan to build AI agents for finance. The founders are young: Forbes said Mateega is 23 and Georgescu is 22. The company doesn't sell ordinary data labeling. It recruits domain experts, including doctors, lawyers, software engineers, and finance specialists, to produce training data and reinforcement-learning environments that capture how professionals work through hard problems. If you've watched large models get better at fluent answers but still stumble on actual judgment, you can see why labs are paying for this. The easy internet text has already been swallowed. The harder material sits in professional habits and edge cases - the steps experienced workers barely think to explain. Singapore's founders now build with Cursor, Lovable and Replit, but ManpowerGroup's 2026 survey found AI development and AI literacy are the country's two hardest skills to hire for. The tools multiplied faster than the engineers who can be trusted to run what they build. - how to fix Singapore developer shortage crisis - AI coding tools versus hiring real engineers AfterQuery said in April that it had surpassed $100 million in annual revenue run rate when it announced the Series A, led by Altos Ventures with participation from The Raine Group, Y Combinator, BoxGroup, and Latitude Capital. Forbes reported this week that Mateega later posted on X that recurring revenue had reached the hundreds of millions. That isn't a small difference. In this market, revenue speed is what gives investors permission to pay prices that look absurd in any normal software cycle. The customer list is why the valuation is easier to understand, even if it is still hard to underwrite from the outside. TechCrunch reported that AfterQuery counts Nvidia and Legora among its customers, along with the Korean AI lab Motif Technologies. Forbes said its work was used in Nvidia's Nemotron models and that the company has also worked with Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati. AfterQuery declined to comment to Forbes. The data race. Look at the neighbors. Mercor, another AI data company founded by young Bay Area founders, raised at a $10 billion valuation in 2025 and was in talks this summer for a round that could value it at $20 billion, according to Forbes and Bloomberg reporting cited by TechCrunch. Scale AI, the older name in this field, became even more central after Meta invested in the company in 2025 in a deal that valued it at $29 billion and brought co-founder Alexandr Wang into Meta's AI effort, according to TechCrunch and Reuters. AfterQuery is running the same race on a shorter clock. Mercor has built around recruiting skilled people to help AI labs improve models. Scale started with annotation and grew into a core supplier for model builders. AfterQuery is trying to push further into expert workflow data, where the point isn't just whether an answer is correct but whether the model can follow the professional path to get there. The market is doing the rest. Frankly, the valuation says as much about buyer pressure as it does about AfterQuery itself. Frontier labs need better data for reasoning, coding, law, medicine, finance, and enterprise work, and they need it now. You can buy more GPUs if Nvidia will sell them to you. You can't instantly manufacture thousands of reliable experts who can turn tacit knowledge into training material a model can learn from. This is the risk. Run-rate revenue can be real and still fragile. A startup can book a huge month from AI labs, annualize it, and look like a rocket, while the durability of those contracts remains hard to judge from the outside. Big customers can squeeze price or switch suppliers. They can build their own internal data teams too, once the process becomes clearer. MIT's Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training released a report on August 25 finding that AI can now credibly complete almost any undergraduate assignment, from essays to proofs to code. The committee is urging departments to fast-track curriculum changes and instructors to shift toward oral exams and portfolios... - ai can now complete undergraduate assignments at mit - how generative ai is changing college coursework standards For now, AfterQuery has the advantage every founder wants: proof that customers are buying before the company has had time to look fully grown. The $3.2 billion price may hold, or it may become another marker of how hot private AI markets became in 2026. Either way, the lesson is plain enough. The new scarce input in AI isn't only compute. It is expert human work, packaged tightly enough that labs can train on it. Elroy is a digital marketer and developer from Goa, with over a decade of experience web development and marketing. He has been associated with several startups and serves currently as an Editor to the Asia Pacific Industrial magazine. He occasionally writes on Startup Fortune about technology and automation.