Introducing NVIDIA Nemotron 3 Ultra. A frontier smart open model built for long-running agents that need to plan, reason, use tools and keep working across complex coding, research and enterprise workflows. Up to 5x faster inference and up to 30% lower cost for agentic tasks. Learn more:

@Astrodevil_ Proudly running on @nebiustf ☝️

Long-running agents need a place to leave their work. Not inside some temporary runtime. Somewhere verifiable, portable, and still there tomorrow.

500B params, great model, you just need a GPU like this to run one :)

NVIDIA: "We made the model 30% cheaper to run!" Also NVIDIA: "Please host it on a minimum of 8x GB200 servers."

Nemotron 3 Ultra (what’s the reasoning behind this name lol) has a massive 550B total parameters. While this scale makes it pretty much unusable on most consumer hardware, its hybrid Mamba-Transformer design keeps it surprisingly efficient for a model of this size with strong long-context performance.

Shipped it in @CommandCodeAI $1 Go plan with 10x credits. $10 and made Nemotron 3 Ultra 2.3x cheaper, so $1 stretches to $20 of usage. Have at it y'all!!

Classic Nvidia: open-sourcing the picks and shovels for agentic AI while locking in the hardware monopoly. Smart capital allocation — they win either way.

Is this your product 😭💀 when

I got early access and broke it down here:

🔥🛢🔥

Nemotron 3 Ultra is the fastest US open weights model out right now — 48 on the Intelligence Index, 550B params but only 55B active. Not quite touching K2.6's 54 yet, but the throughput is insane at 300+ tok/s. Available in NVFP4 too so it'll run on decent hardware. This is the kind of model that makes local inference actually competitive with the frontier.

Excited about NVIDIA Nemotron 3 Ultra: an open, frontier class model designed for long-running agents requiring planning, reasoning, and tool use across complex coding, research, and enterprise workflows. Up to 5x faster inference and 30% lower cost for agentic tasks. #Cybernorse

GLM said "we're competitive" then posted a 40% on agent productivity 💀

this model is post trained specifically with agentic workflows in mind and it is very impressive. i went through the paper properly and found out the most interesting stuff from it. complete breakdown here:

Most handsome and having good physique CEO😂😂 this means much richer than Musk🦅

bruh, brilliant–––but who is naming these? xD

1M context window for long-running agents is a genuinely interesting constraint to work with. The question for teams building on this: what happens to all that context data? Prompts, tool outputs, intermediate reasoning - at enterprise scale that's a significant data handling problem. Model is on Hugging Face:

When long horizon planning meets short term drrinking?

Solid move from NVIDIA, open models with real agent capability is where the value is

A cheaper model that skips verification is just moving the bill to a human at 2am.

5x faster inference, but don't forget to budget for those 8x GB200 servers. who knew speed came with a price tag?

Hell yeah, tech boosting customs to shield American jobs and stop unfair imports. America First hitting different!

Thank you @NousResearch and @NVIDIAAI for making this free, I’ve loaded it and running it on my agent!

A 300B range model with 30B activations is an explored territory by American OSS AI. A combo of intelligence with speed. A good place for Nvidia to target its hardware for medium sized customers.

Teu pai tá trabalhando @pedropverani

The interesting bit is not 550B by itself. It is open weights aimed at tool use, planning, and long jobs where reliability starts eating the budget.

Lolol Nvidia are masters at releasing LLM's that run AMAZING! If you buy $50K of their GPU's

Excited to see NVIDIA Nemotron 3 Ultra's powerful open agentic capabilities its long-running reasoning, tool use, and cost-efficient inference will be a game-changer for building advanced decentralized AI agents that leverage @Filecoin's reliable, censorship-resistant storage

open models are all the same change my mind

now if only you could sell a high vram consumer hardware piece...

Thank you, Jensen. This is what the West needs as we can see China is leading with the open weight models which gives a lot of freedom to people to either run them locally or on a Cloud meanwhile the likes of anthropic and OpenAI provide you decent models but completely opaque and give you no flexibility whether to host them locally and they like to lock you in to their platforms which is not the way of the future no matter how much money Anthropic makes or OpenAi makes. I don’t think that they are the future. Yes they have had some breakthrough but I think what you’re building is bigger and better.

NVIDIA Nemotron 3 Ultra shows how quickly agentic AI is evolving. Faster inference, lower costs, and stronger reasoning capabilities are exactly what's needed to power the next generation of autonomous AI workflows.

Cool. Now question out of context: Will your next GPUs be affordable or I need to take a mortgage ?

@nvidia sounds like a beast for coders. Can't wait to see how it handles real-world tasks.