Nemotron-3-Super-Q4_K_M-YaRN-1M-GGUF

Nemotron-3-Super — Q4_K_M GGUF with YaRN 1M (Ollama / llama.cpp)

Community llama.cpp / Ollama pack derived from the Ollama library nemotron-3-super Q4_K_M weights (NVIDIA Nemotron 3 Super, ~120B total / ~12B active MoE).

This is not an official NVIDIA Transformers / NIM / NVFP4 safetensors checkpoint. Official weights live under nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 (and sibling FP8/BF16 repos). This repo is a Q4_K_M GGUF with metadata-only YaRN so runners that honor n_ctx_train accept 1,048,576 tokens.

License: NVIDIA Nemotron Open Model License (same family as the official Super card). The LICENSE file is the NVIDIA Open Model License Agreement text (Last Modified: October 24, 2025). See also license link.

---

What you get

The single ~87 GiB file exceeds Hugging Face’s 50 GB per-file limit, so this pack is llama-gguf-split --split-max-size 40G. Ollama / llama.cpp load the set from the first shard path.

1M context is not extra trained weights. Tensors are unchanged from the Ollama Q4_K_M blob. Only GGUF key-values were rewritten:

Sparse attention keeps 1M KV cheap on unified-memory hosts (~8 GB KV class + ~81–87 GB weights ≈ ~94–95 GB resident on a DGX Spark when fully allocated).

---

Why this pack exists

NVIDIA’s Super card targets up to 1M context. The published Ollama library GGUF only advertised 262144, and stock Ollama caps num_ctx at GGUF n_ctx_train. Setting PARAMETER num_ctx 1048576 alone left ollama show at 262K.

This pack patches KV (same pattern as our Qwen3.8 NVFP4+VL YaRN pack) so Ollama / llama.cpp will allocate the full window.

---

Validated on DGX Spark (GB10, 128 GB unified)

Host: Docker ollama/ollama (~0.32.x), tag nemotron-3-super:1m, Spark OLLAMA_CONTEXT_LENGTH left at 65536 (fallback only — Modelfile pins 1M).

Secret used for the full-depth run: SECRET_CODE_NEMO_SUPER_1M_FULL_9917 (recalled correctly). Global env was not raised to 1M; precedence is request → Modelfile → OLLAMA_CONTEXT_LENGTH.

---

Use

Needs a recent Ollama / llama.cpp build that understands nemotron_h_moe. Keep RENDERER/PARSER nemotron-3-nano for tools + thinking.

Ollama (after hf download)

hf download mixbits/Nemotron-3-Super-Q4_K_M-YaRN-1M-GGUF --local-dir ./nemotron-3-super-yarn1m
cd ./nemotron-3-super-yarn1m
# Modelfile FROM points at shard 00001-of-00003; keep all three shards together
ollama create nemotron-3-super:1m -f Modelfile
ollama run nemotron-3-super:1m

API smoke

curl -s localhost:11434/api/chat -d '{
  "model": "nemotron-3-super:1m",
  "stream": false,
  "think": false,
  "options": {"num_predict": 64},
  "messages": [{"role": "user", "content": "Hello"}]
}'

Leaving Open WebUI num_ctx as Default uses Modelfile 1M. Do not raise a global Ollama OLLAMA_CONTEXT_LENGTH to 1M just for this tag — that hits every unpinned model.

Recreate YaRN GGUF from stock Q4

If you still have the library blob (or re-ollama pull nemotron-3-super:latest):

python scripts/nemotron-super-yarn-1m-rewrite.py \
  /path/to/stock-q4.gguf \
  ./Nemotron-3-Super-Q4_K_M-yarn1m.gguf

---

Honesty checklist

• Do not claim official NVIDIA NVFP4 / FP8 / BF16 safetensors.
• Do not claim we re-quantized from Hub weights; language tensors came from the Ollama library Q4_K_M blob, then KV rewrite only.
• Do not claim Modelfile-only num_ctx is enough without GGUF n_ctx_train ≥ 1M.
• Do claim DGX Spark validation: 1M KV alloc + needle through ~512k and a single-shot ~980k recall PASS (2026-08-23).

---

Attribution

• Model family: NVIDIA Nemotron 3 Super (tech report, Hub NVFP4)
• Q4_K_M distribution: ollama.com/library/nemotron-3-super
• YaRN 1M GGUF rewrite + Spark validation: mixbits (this repo)