Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(outreach): draft the questions three.ws puts to the Nemotron Labs live panel

nirholas0dc6bf546102
diff --git a/docs/nvidia-nemotron-ask-the-experts-questions.md b/docs/nvidia-nemotron-ask-the-experts-questions.md
new file mode 100644
index 000000000..fb3180a80
--- /dev/null
+++ b/docs/nvidia-nemotron-ask-the-experts-questions.md
@@ -0,0 +1,113 @@
+# Ask the Experts: Nemotron Open Family (Nemotron Labs live on X)
+
+Questions for three.ws to drop into the live chat of NVIDIA's "What's New in the
+Nemotron Open Family" stream (@NVIDIAAI). Ordered by what actually helps this platform
+and by how likely a panel is to read it out.
+
+## What the panel is fielding
+
+Public state of the family as of this stream:
+
+| Model | Params (total / active) | Notes |
+|---|---|---|
+| Nemotron 3 Nano | 31.6B / 3.6B | 1M token context, 24.6 GB, consumer-GPU class |
+| Nemotron 3 Super | 120.6B / 12.7B | GTC, March 2026 |
+| Nemotron 3 Ultra | 550B / 55B | June 2026 |
+| Nemotron 4 | ~1T, in training | reported for late 2026 |
+
+The family ships open weights, open training data, and open recipes (roughly 10T tokens
+of released datasets), with a hybrid mixture-of-experts architecture aimed at agentic
+work. Every question below is written against that, not against generic "what is an LLM"
+framing, because the read-out-loud questions are the specific ones.
+
+## Rules for posting into a live chat
+
+- One question per message. Two questions in one line get half-answered.
+- Under about 200 characters. Long messages are skipped on air.
+- Lead with the constraint, not with our product. "3.6B active on an L4" is a hook;
+ "we built a 3D avatar platform" is an ad and gets skipped.
+- Post from the three.ws handle, ask the follow-up only after they answer the first.
+
+## The questions
+
+### 1. Interactive latency next to a render loop (best single ask)
+
+> Nemotron 3 Nano at 3.6B active: what time-to-first-token are you seeing on an L4 when
+> the GPU is also feeding a real-time render loop? Any guidance on sharing one GPU
+> between inference and graphics?
+
+Why it lands: it is a hardware question with a number in it, it is the exact thing our
+GPU workers do, and nobody else in the chat will ask it. This is the one to post first.
+
+### 2. Tool calling under an agentic harness
+
+> For agentic use: how does Nemotron 3 hold up on strict JSON schema adherence over long
+> multi-tool runs, and is there a reasoning-budget control to keep tool latency bounded?
+
+Why it lands: "agentic AI" is the family's own positioning, so they have prepared
+material, and schema adherence is what actually decides whether an agent runtime can use
+a model in production.
+
+### 3. What 1M context costs in memory
+
+> The 1M context on Nano is the headline. What does the KV cache cost at 1M on a single
+> 24 GB card, and what retrieval accuracy do you measure at full length vs 128k?
+
+Why it lands: it is a friendly but real question. Long-context claims usually degrade,
+and asking for the honest curve reads as a builder, not a critic.
+
+### 4. Fine-tuning on a narrow domain without losing the agentic behavior
+
+> With the open recipes out there: what is the recommended way to fine-tune Nano on a
+> narrow domain corpus without degrading the tool-calling behavior you post-trained in?
+
+Why it lands: this is the question every team that adopts an open model hits in week
+two, and the open-recipe release is exactly what makes it answerable on air.
+
+### 5. Quantization floor for consumer hardware
+
+> Where does quality actually break down on Nano: FP8, NVFP4, lower? Which quantization
+> do you ship as the recommended default for consumer GPUs?
+
+Why it lands: it converts the "24.6 GB fits consumer hardware" claim into something a
+builder can plan capacity against.
+
+### 6. Vision and 3D
+
+> Anything in the Nemotron VL line aimed at 3D assets, so a model can look at a rendered
+> mesh and judge topology or texture quality? Curious how it lines up with Cosmos.
+
+Why it lands: this is our lane (text to 3D, rigging, avatar QA), and it invites them to
+talk about physical AI, which they like talking about.
+
+### 7. Serving stack on day one
+
+> Which serving path do you consider first-class for Nemotron 3: TensorRT-LLM, vLLM, or
+> SGLang? Where does the hybrid MoE architecture need a version bump to run right?
+
+Why it lands: hybrid architectures routinely land ahead of runtime support, and knowing
+which runtime is blessed saves a deployment week.
+
+### 8. What stays open in Nemotron 4
+
+> With Nemotron 4 reported at trillion scale: does the same openness bar hold there,
+> weights plus data plus recipes, or does the tier change with size?
+
+Why it lands: it is the forward-looking question of the stream, and it is a fair one to
+ask a team whose whole pitch is openness. Post it late, near the end.
+
+### 9. The accuracy follow-up
+
+The chat already asked "how do you increase the accuracy?" and that question is too
+broad to get a useful answer. A sharper version, posted as a follow-up:
+
+> Following the accuracy question: for a narrow domain, what moves the needle more in
+> your experience, more post-training data or more reasoning tokens at inference?
+
+Why it lands: it rescues a question the panel already saw, which makes it likely to be
+picked up, and the answer is genuinely useful to us.
+
+## If we only get one message in
+
+Post question 1. It is specific, it has numbers, it is a hardware question at a hardware
+company, and the answer directly informs how we size the GPU workers.