NVIDIA is developing a new AI model family called Nemotron 4, with the explicit goal of competing against the world's leading open-source models, according to CNA, which cited a report by U.S. tech outlet The Information based on people involved in the project.
The strategic backdrop matters. CNA reported that NVIDIA is one of the few major U.S. companies to release open-source models at all, and that as AI development costs have climbed, affordable Chinese models have moved closer in performance to the leading systems from Anthropic and OpenAI — a dynamic that has drawn renewed attention to open-source AI this year. Inside.com.tw, also citing The Information, added more texture: with AI training and inference costs described as surging this year, low-cost Chinese models such as Qwen (千問) and DeepSeek (深度求索) have narrowed the performance gap with top Anthropic and OpenAI systems, pulling market attention back toward the open-source route. The same report noted that a string of autonomous AI agent intrusion incidents has simultaneously raised concerns about the lack of usage restrictions on open-source models in security-sensitive applications.
NVIDIA's generative AI vice president, Kari Briski, has framed the company's stance directly, saying that "every enterprise, every country should have access to advanced open-source models," according to Inside.com.tw. Despite the reporting, CNA noted that NVIDIA had not responded to Reuters' request for comment.
According to CNA, multiple employees involved in the project said the largest Nemotron 4 model is expected to carry at least 1 trillion parameters. Inside.com.tw, citing The Information report published Tuesday, August 11, added a direct comparison: the largest version of Nemotron 4 is expected to have at least 1 trillion parameters — roughly twice the size of the current flagship, Nemotron 3 Ultra.
| Model | Parameter scale | Key milestone | Status |
|---|---|---|---|
| Nemotron 3 family | Not disclosed | Late 2025 | Launched, as Chinese AI labs' products proliferated (CNA) |
| Nemotron 3 Ultra | Baseline (current flagship) | — | Currently in production (Inside.com.tw) |
| Nemotron 3.5 Lightning | Not disclosed | Aug. 11, 2026 | Launched (CNA) |
| Nemotron 4 (largest model) | At least 1 trillion parameters, ~2x Nemotron 3 Ultra | Earliest target: late fall | Final training incomplete, no release date set (CNA; Inside.com.tw) |
CNA reported that NVIDIA has not set a release date for Nemotron 4 and has not finished final training, but that employees said the model could be ready as early as late fall. This lines up with Inside.com.tw's account, which cited NVIDIA staff estimating the model could be battle-ready at the earliest in the deep fall of this year, while noting the release date remains undetermined.
The project follows a clear cadence: CNA reported that late last year, as products from Chinese AI labs proliferated rapidly, NVIDIA launched its third-generation open-source model family — the lineage that includes Nemotron 3 Ultra, the current flagship Nemotron 4 is being measured against.
Nemotron 4 is not arriving in isolation. CNA reported that NVIDIA also launched Nemotron 3.5 Lightning the same day (Aug. 11), a new addition to its product lineup built for tasks including code review, tool use, security alert monitoring, and billing-inquiry responses. Alongside it, NVIDIA released NeMo Switchyard, described by CNA as an open-source model-routing library that automatically directs AI tasks to the most suitable model.
These product moves follow broader positioning by NVIDIA on open models. CNA reported that last month, NVIDIA joined an alliance with other companies to develop and share AI safety and security tools, and separately signed a public letter with Microsoft (微軟) and other tech giants expressing support for open-weight models, framed as a way to prevent innovation resources from flowing overseas.
Inside.com.tw, in its own analysis citing The Information, identified two variables facing Nemotron 4. First, a parameter count of 1 trillion is not an absolute lead within the open-source field, and NVIDIA will need to rely on compression techniques and TensorRT optimization to compensate. Second, NVIDIA faces what the outlet described as a subtle competitive-and-cooperative dynamic with customers such as Microsoft and OpenAI at the model layer — companies that are simultaneously NVIDIA's hardware customers and, with a competing model family, potential rivals.
The same report characterized a successful late-fall launch as a potential turning point for NVIDIA: a shift from being a "pick-and-shovel seller" of AI hardware toward a full-stack player combining compute and models.
The reporting from CNA and Inside.com.tw, both citing The Information, converges on the same core numbers and timeline: a model targeting at least 1 trillion parameters — about double Nemotron 3 Ultra — with an earliest possible readiness in late fall, though NVIDIA has confirmed neither the parameter count nor a launch date, and did not respond to Reuters' request for comment. That gap between reported ambition and confirmed fact sits alongside a second tension flagged by Inside.com.tw: NVIDIA co-signed a public letter with Microsoft supporting open-weight models even as the same report describes NVIDIA and Microsoft's customer, OpenAI, as forming a "subtle competitive-and-cooperative" relationship with NVIDIA at the model layer. NVIDIA's parallel launches of Nemotron 3.5 Lightning and the NeMo Switchyard routing library, alongside last month's AI-safety alliance, indicate the company is building out its open-source model ecosystem incrementally while the flagship Nemotron 4 remains unfinished.
According to CNA and Inside.com.tw, both citing The Information, the largest Nemotron 4 model is expected to carry at least 1 trillion parameters — about twice the size of NVIDIA's current flagship, Nemotron 3 Ultra.
CNA reported that NVIDIA has not set a release date and has not completed final training, though employees said the model could be ready as early as late fall. Inside.com.tw similarly cited staff estimating the deep fall as the earliest readiness window.
No. CNA reported that NVIDIA had not responded to Reuters' request for comment on the Nemotron 4 project.
CNA reported that on August 11, NVIDIA launched Nemotron 3.5 Lightning, aimed at tasks like code review and security alert monitoring, and released NeMo Switchyard, an open-source model-routing library.
According to CNYES and Inside, DeepSeek's API documentation quietly switched to DeepSeek-V4-Pro-0813 on August 12, adding a 1,000,000-token context and 384,000-token output. Unofficial scores shared in DeepSeek's community group reportedly put agent performance close to Fable 5, but DeepSeek has not published these results officially.
Google launched the Pixel 11 series on August 12 with its self-developed Tensor G6 chip, raising U.S. prices by $100 across the line to $899–$1,899 and Taiwan prices to NT$29,990–NT$42,990, according to CNA. Device chief Rick Osterloh told CNA that memory shortages are forcing 'every company' to raise prices, and Anue reported TrendForce data showing DRAM contract prices jumped up to 93% in Q1 2026.
According to VentureBeat, SpaceXAI launched an early beta of Grok Bot on August 11, 2026 — an AI agent built to keep executing work inside a user's tools rather than just answering prompts. Grok Bot ships with Cursor Premium Teams at $120 per seat monthly, is bundled into the $200 Cursor Ultra plan, and reaches existing $300 SuperGrok Heavy subscribers too, VentureBeat and iThome report.