NVIDIA-Nemotron-Labs-Teacher-STEM
One-line: NVIDIA’s STEM/general-reasoning teacher from the Nemotron 3 Ultra family (08/14 drop) — 550B/55B LatentMoE hybrid covering math, code, natural sciences, and tool use (code execution + search), positioned vs DeepSeek V4 Pro (High). ⚠ large — datacenter-class (4xGB200/8xH100), not locally runnable. ⚠ no published benchmarks — agent-loop suitability unknown.
⚠ No benchmark numbers published on the card. The accuracy plot (
accuracy_plot.png) is a broken link (404 on the repo), there is no.eval_results/dir. The only quantitative claim is narrative: “matching or outperforming DeepSeek V4 Pro (High) on benchmarks such as GPQA, MMLU-Pro, LiveCodeBench v6, IMOAnswerBench, and Apex Shortlist” — lab-reported, no table, no numbers. Coding/agent cells below are—.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published (narrative: “vs DeepSeek V4 Pro (High)” only) |
| Aider polyglot | — | not published |
| SWE-bench Verified | — | not published |
| SWE-bench Multilingual | — | not published |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published |
| τ-Bench | — | not published |
| ToolACE | — | not published |
| GAIA | — | not published |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published (narrative claim) |
| GPQA Diamond | — | not published (narrative claim) |
| MATH | — | not published |
| AIME 2025 | — | not published |
Efficiency
- Active params: 55B active (550B total, 512 experts/layer, top-22)
- Context length: 262,144 tokens (config
max_position_embeddings) - VRAM (fp16 rough): ~1.1 TB full-model (datacenter only; min 4xGB200 / 8xH100 per card) — ⚠ needs quantization (NVFP4 recipe) and multi-GPU even then
What makes it notable
Second of five Nemotron 3 Ultra Teacher checkpoints released 08/14/2026 (siblings: Competition-Coding, Chat, Instruction-Following, General-Reasoning). Produced from the post-trained Ultra student (550B-A55B, LatentMoE hybrid Mamba-2+MoE+Attention, MTP, OpenMDW-1.1) plus reasoning-focused SFT+RL. Unlike the Competition-Coding teacher it’s a general reasoning model — but the training data explicitly covers code and tool use (code execution + search), which is why it’s relevant to coding/agent pipelines. NVIDIA’s headline positioning: matches or beats DeepSeek V4 Pro (High) on GPQA, MMLU-Pro, LiveCodeBench v6, IMOAnswerBench, Apex Shortlist — unverifiable from the card itself. Custom architecture (NemotronHForCausalLM) — needs the repo’s model code; no llama.cpp/Ollama path. No independent evals exist yet; treat all claims as lab-reported until verified.