Sources: Nvidia is developing a Nemotron 4 model with 1T+ parameters, up from Nemotron 3 Ultra's 550B parameters but smaller than leading Chinese open models

The Information:

The Information 2026-08-12

Context & Ripple Effects

Nvidia’s Nemotron line has moved quickly from a 30B-to-~500B hybrid MoE family to a 120B Super release, alongside a stated plan to fund open-model development. Its 550B Ultra model was described as the strongest open U.S. model but still behind Kimi K2.6, making the reported next step chiefly a response to an open-model performance gap.

The roadmap is also broadening rather than simply scaling up: Nvidia released a 30B Lightning model and an open routing library while pursuing a much larger model. That pairs smaller deployable models with a prospective flagship for buyers choosing among open-weight systems.

First-order effects

Nvidia is raising the upper end of its open-model roadmap beyond Nemotron 3 Ultra’s 550B scale, while retaining a stated size gap with leading Chinese open models.
Nemotron users and prospective buyers gain a clearer signal that Nvidia intends to compete at the trillion-parameter tier, not only through its existing mid- and large-scale releases.

Second-order effects

Chinese open-model developers retain a performance-and-scale benchmark that Nvidia is explicitly chasing, increasing pressure on Nvidia to translate parameter growth into a stronger competitive result than Ultra delivered.
Enterprises evaluating Nvidia’s 30B Lightning, 550B Ultra, and a future flagship will face a more segmented model-selection process, elevating the value of routing tools such as NeMo Switchyard.

Third-order effects

If Nvidia sustains the progression from Nemotron 3 Super through larger open releases, open-weight competition will increasingly be organized around full model portfolios—small, routable models alongside frontier-scale flagships—rather than a single benchmark model.
The reported Chinese scale lead suggests that parameter count is becoming a visible competitive signal in the open-model market, even as model buyers will still need to judge the resulting capability and deployment economics.

The trend: Open-model competition is shifting toward portfolio strategies that combine lightweight deployment models, routing software, and frontier-scale flagships.

Discussion

@andrewcurran_ Andrew Curran on x
NVIDIA is building its next-gen Nemotron 4 family to compete directly with leading Chinese open models and secure the open-weight crown for the U.S. The largest version will have at least 1 trillion parameters, according to original reporting from The Information. [image]
@scaling01 @scaling01 on x
“The largest Nemotron 4 model is expected to have at least 1 trillion parameters” lets fugging go, we need more huge open-weight models
@paulharper.eurosky.social Paul Harper on bluesky
Going from 550B parameters to 1T+ parameters is a significant doubling of my lack of interest... [embedded post]
@aravsrinivas Aravind Srinivas on x
A great American open weights MoE model that can run efficiently on your laptop or local hardware like the DGX Spark! You can use the larger Nemotron Ultra on Perplexity!
@miaai_lab Mia on x
Nvidia's Nemotron 3.5 Lightning is insanly fast on a DGX Spark, I'm blown away 🤯 [image]
@deryatr_ Derya Unutmaz on x
Very excited that, after Meta released its open-source AI models yesterday, this morning another major American open-weights model was released by NVIDIA! Nemotron 3.5 lightning is a 30B-parameter model, 4x faster than similar sizes, with open weights, making it highly
@fastinoai @fastinoai on x
In collaboration with @nvidia we're releasing two new open weight models: Fastino-Nemotron-3.5-Lightning- Finance and Fastino-Nemotron-3.5-Lightning- Healthcare. Working closely with the Nemotron team, we developed both models on Nemotron 3.5 Lightning using the Fastino [image]
@pidotdev @pidotdev on x
Nvidia Nemotron 3.5 Lightning delivers leading accuracy on agentic coding tasks and up to 4x the speed of comparable open models. With open weights, datasets, and recipes, Nemotron is open and easy to customize. Welcome to Pi, Nemotron 3.5 Lightning! [image]
@liangsays Brent Liang on x
thank you @nvidia for including @MTSlive under embargo with early access we've been playing with nemotron 3.5 lightning. very excited it's now out! [image]
@jensenhuang Jensen Huang on x
Lightning strikes for continuous and long-run agents! Nemotron 3.5 Lightning is smart, fast, efficient and open.
@kimmonismus @kimmonismus on x
One day after Meta released Muse Glimmer, NVIDIA launched Nemotron 3.5 Lightning and NeMo Switchyard, showing a different approach to local agentic AI. Really cool. Lightning is a 30B MoE with only 3B active parameters, distilled from Nemotron 3 Ultra and designed for the [image]
@lmstudio @lmstudio on x
Nemotron 3.5 Lightning is available in LM Studio! The model is 30B MoE (3B active), can run very fast, and is trained for high volume agentic use cases. Model page: https://lmstudio.ai/...
@nvidiaai @nvidiaai on x
Long-running agents spend most of their time executing: calling tools, validating results and delegating work. Nemotron 3.5 Lightning is built for this high-volume execution, at a size that can run anywhere from an NVIDIA DGX Spark to the data center. See it running agentic [vide…
@0xsero @0xsero on x
I've been testing this model, excellent agent. With DSpark it's INCREDIBLY fast on a DGX Sparks 240+ tok/s Highly recommend trying this out. Not the best for coding but it's great at tool calling, exploration, apps, and being a general assistant. Best Nvidia model to date
@openrouter @openrouter on x
NVIDIA Nemotron 3.5 Lightning is now live on OpenRouter. A 30B hybrid MoE with 3B active params, distilled from Nemotron 3 Ultra. Built for high-volume, specialized AI agent workloads, delivering up to 4× higher throughput and up to 30% faster task completion compared to simila…
@trajectorylabs @trajectorylabs on x
Continual learning is a bet that the retraining loop will get cheaper over time. With larger models, you can maybe run this loop once every few weeks. But with smaller models, you can run it nightly, per customer. And it keeps recursing: a model per company, then a model per [ima…
@llm_wizard Chris on x
🌩️Nemotron 3.5 Lightning is the newest member of the Nemotron 3 family - and the whole thing is that it's fassssst (and pretty smart too) - it's a 30B A3B model - runs on DGX Spark, ollama/llama.cpp support and all the good stuff. AND YOU KNOW IT'S AN OPEN LICENSE. Pushing [image…
@artificialanlys @artificialanlys on x
NVIDIA has just released the first Nemotron 3.5 model: Nemotron 3.5 Lightning, a highly efficient small open weights model with performance similar to gpt-oss-120b at around a quarter of the total parameters Key takeaways...Major intelligence jump...Optimized for efficiency...Mea…
@nvidiaai @nvidiaai on x
Lightning is built to specialize. Post-train Nemotron 3.5 Lightning with NVIDIA NeMo for your domain data, tools, workflows and policies. Across cybersecurity, coding, legal and energy tasks, post-training improves accuracy for specialized work. [image]
@nvidiaai @nvidiaai on x
Lightning pairs strong accuracy with speed. On PinchBench, it reaches 86% accuracy while completing 10,000 tasks 35% faster than Qwen3.6 35B at similar accuracy. [image]
@nvidiaai @nvidiaai on x
Introducing NVIDIA Nemotron 3.5 Lightning⚡ An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models. [image]

Chronicles