NVIDIA dropped Nemotron 3 Super today, a 120 billion parameter open-source reasoning model built specifically for agentic AI workloads. And in the same news cycle, Anthropic shipped major upgrades to Claude in Excel and PowerPoint, giving enterprise users a single AI session that moves between spreadsheets and slides without losing context. Two very different companies. One very clear direction. AI is getting embedded everywhere and the open-source side of the race just got a serious contender.

Most people think of NVIDIA as the chip company that everyone else depends on. Today they reminded the industry they are also a serious model company.

Nemotron 3 Super is a 120 billion total, 12 billion active parameter model built on a hybrid Mamba-Transformer mixture-of-experts architecture. That gap between total and active parameters is the whole point, the model activates only what it needs, which is how NVIDIA delivers over 5x higher throughput than the previous Nemotron Super while maintaining frontier-level accuracy.

Here is what makes this release different from everything else in open source right now:

The architecture. The hybrid Mamba-MoE design delivers 4x memory efficiency over traditional models. Mamba layers handle the memory-heavy lifting while Transformer layers kick in for high-accuracy reasoning. The result is a model that does not force you to choose between speed and intelligence.

The context window. 1 million tokens natively. For agentic systems that are constantly re-sending history, tool outputs, and reasoning steps, context explosion is one of the biggest unsolved problems. Nemotron 3 Super addresses it directly.

The reasoning budget. Developers can dial exactly how hard the model thinks. Full reasoning mode for complex multi-step problems. A capped budget for latency-sensitive applications. This kind of granular control at the API level is rare even in closed models.

The openness. This is not just open weights. NVIDIA is releasing the training datasets, reinforcement learning environments, and post-training recipes. You can inspect what the model learned on, reproduce its training, and customize it for your domain. No other major Western lab is doing this at this scale.

On benchmarks, Nemotron 3 Super scored 85.6% on PinchBench, the standard suite for agentic workloads, surpassing Claude Opus 4.5, Kimi 2.5, and GPT-OSS 120B.

It runs today on Hugging Face, vLLM, SGLang, Ollama, and llama.cpp. Available on DGX Spark, H100, and B200 GPUs.

The AI industry has spent two years treating open source as the scrappy underdog. DeepSeek made that narrative complicated. Nemotron 3 Super makes it obsolete.

NVIDIA is not just releasing weights. They are releasing the entire stack, datasets, RL environments, training recipes, under a permissive license that allows commercial deployment without crediting NVIDIA. They are actively trying to make the open-source ecosystem win because a thriving open ecosystem sells more GPUs.

⚡ The Vibe Check: Jensen Huang has been saying "open innovation is the foundation of AI progress" for months. Today he backed it up with a model that beats closed frontier models on agentic benchmarks. The gap between open and closed is closing faster than anyone expected.

Some Cool Stuff From Anthropic

If you use Excel or PowerPoint professionally, today's update from Anthropic is the one that actually changes your workflow starting tomorrow morning.

Claude for Excel and Claude for PowerPoint now share full context across your entire open session. That sounds small until you think about what it means in practice.

Picture this. You are staring at a financial model, FY2026 through FY2028 projected out, and something feels off about the margin assumptions. You type into the Claude sidebar: "uhh, that profit margin assumption looks kinda too high. What did Acme actually say on the call?" Claude goes and finds it. Then you type /scenario-analysis and tell it to use the conservative margin as base case, full guidance as best case, competitive pressure compression as worst case, and add the Scenarios sheet to the model. It does all of it. In the same session. Without you touching a formula.

That is what one continuous session across files actually feels like in practice. One session. All context carried forward. No re-prompting.

The other major addition is Skills: reusable one-click workflows that turn your team's best practices into repeatable actions inside the sidebar. Anthropic shipped a preloaded starter set covering the most common use cases. For Excel: auditing models for formula errors and balance sheet integrity, building LBO, DCF, and 3-statement model templates, running comparable company analyses, and cleaning messy spreadsheet data. For PowerPoint: building competitive landscape decks, updating existing decks with new data, and reviewing investment banking decks for number consistency and language polish.

Every skill you have already set up in Claude on desktop or web works inside the add-ins automatically. And Instructions let teams set persistent preferences, always use the firm's number formatting, keep PowerPoint bullets to one line, flag hardcoded assumptions, set once, applied every time without additional prompting.

On deployment, Claude for Excel and PowerPoint now route through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Enterprises already locked into a cloud compliance setup do not need to change anything.

Available today in beta for all paid plan users on Mac and Windows.

This is a single mechanical keyboard keycap with the Anthropic logo on it. One key. One purpose. You press it and presumably Claude does everything else.

Honestly at this point that is not even a joke. Claude is inside your Excel. Inside your PowerPoint. Inside your browser. Inside your terminal. Inside your Slack. The keycap is just being honest about where we are headed. Why have a full keyboard when one button is doing 90% of the work anyway.

The people who made this understood the assignment. The rest of us are still pretending we need the other 103 keys.

Someone sell this. We will be the first customer.

NVIDIA shipped an open-source model that beats closed frontier models on agentic benchmarks and gave the entire developer community the training stack to prove it. Anthropic quietly made the case that AI's real enterprise battleground is not the chat interface, it is the spreadsheet and the slide deck where the actual work happens. The skills feature in particular is interesting because it turns individual AI workflows into organizational assets. One analyst figures out the right way to run a variance analysis. Save it as a skill. Now the whole team runs it in one click. That is how AI goes from a personal productivity tool to an institutional one. The pace right now is relentless. The only question is whether you are building with these tools or watching someone else do it first.

Stay building. 🤖