Local AI Crossed the “o3” Barrier, Nemotron 3 Super Solves the Model-Killer Puzzle

The “Teacher’s Word” logic puzzle became a benchmark challenge for local LLMs after a viral Reddit thread claimed that not a single local model could solve it consistently.

At the time, OpenAI’s o3 appeared to dominate this category of multi-agent epistemic reasoning.

This video documents one of the first successful local runs using NVIDIA Nemotron-3-Super-120B on a single-node DGX Spark setup.

Interestingly, NVIDIA’s hosted online demo initially failed with an internal server error, while the fully local setup completed the reasoning successfully.

CHALLENGE (Try it yourself on local AI):

A teacher writes six words:
“cat dog has max dim tag.”

She gives Albert, Bernard, and Cheryl each one letter from the same word.

She asks Albert if he knows the word.
He immediately says YES.

She asks Bernard.
He thinks, then says YES.

She asks Cheryl.
She thinks, then says YES.

What is the word?

SETUP:

• Model: Nemotron-3-Super-120B-A12B
• Architecture: Hybrid LatentMoE
• Hardware: NVIDIA DGX Spark
• Interface: Open WebUI / local inference
• System Prompt: None
• Hyperparameters: Default
• Result: Correct deduction of “DOG”
• Video: Unedited screen recording

Why this matters:

Most local models fail at tracking recursive “states of knowledge” reasoning and hallucinate the deduction chain.

This run demonstrated that frontier-level reasoning was becoming possible on sovereign local infrastructure, narrowing the gap between cloud-only systems and local AI.

Resources and reproducible setup are pinned in the comments.

Disclaimer:

Results vary depending on quantisation, inference engine, runtime configuration, prompting strategy, concurrency settings, and hardware optimisation.

Видео Local AI Crossed the “o3” Barrier, Nemotron 3 Super Solves the Model-Killer Puzzle канала Rajendra Rawat