model-pull

name	model-pull
description	Pull local inference models (Hermes V4, Qwen3 embed, Gemma 4 multimodal, NeMo Omni, Unsloth GGUF variants) via Ollama or HuggingFace CLI. Node-aware routing: ≥70B → SPARK only; 8B → any node with Ollama; embedding models → Z890/RDNA4/Spark; multimodal (text/image/audio in, text out) → Spark/5090/4090; NeMo Omni → Spark only. Use /model-pull <model-id> or describe what you want to pull.
disable-model-invocation	true

model-pull

Pull models from the Phase D catalog into local Ollama or HuggingFace cache.

License policy (verified on HF, 2026-07)

Every model in this table is license-verified on the Hugging Face Hub and open for commercial use (Apache-2.0 / MIT / permissive) per the open-source-only policy. Two rows are flagged ⚠ verify — their upstream licenses (Llama Community, NVIDIA) are not the permissive class and must be license-checked against the policy before a production pull.

Legacy Gemma retired. Gemma 2 / Gemma 3 / embeddinggemma-300m carry license:gemma (Google Gemma Terms of Use — gated, prohibited-use policy) and are NOT open-source-policy compliant. Gemma 4 (E2B/E4B/31B/26B-A4B, released 2026, license:apache-2.0, ungated) is the compliant successor and the only Gemma family this skill pulls.

Models

Flare ID	Ollama tag	HuggingFace ID	HF license	Node
pmoves/qwen3-embed-0.6b	n/a	Qwen/Qwen3-Embedding-0.6B	apache-2.0	z890, rdna4, 4090
pmoves/qwen3-embed-4b	n/a	Qwen/Qwen3-Embedding-4B	apache-2.0	dgx-spark, 5090
pmoves/gemma4-e4b	gemma4:e4b	google/gemma-4-E4B-it	apache-2.0	dgx-spark, 5090, 4090
pmoves/gemma4-31b	gemma4:31b	google/gemma-4-31B-it	apache-2.0	dgx-spark only
pmoves/qwen3-coder-30b	n/a	Qwen/Qwen3-Coder-30B-A3B-Instruct	apache-2.0	dgx-spark, 5090
pmoves/hermes-v4-8b	hermes3:8b	NousResearch/Hermes-3-Llama-3.1-8B	⚠ verify (Llama 3.1 Community)	dgx-spark, 5090, 4090, rdna4
pmoves/hermes-v4-70b	hermes3:70b	NousResearch/Hermes-3-Llama-3.1-70B	⚠ verify (Llama 3.1 Community)	dgx-spark only
pmoves/nemo-omni-vl	n/a	nvidia/NVLM-D-72B	⚠ verify (NVIDIA — may be CC-BY-NC)	dgx-spark only

Unsloth variants (GGUF, optimized for local inference):

• unsloth/Hermes-3-Llama-3.1-8B — 4090/5090 friendly (⚠ inherits Llama 3.1 Community license)
• unsloth/Hermes-3-Llama-3.1-70B — Spark only (⚠ inherits Llama 3.1 Community license)

Commands

HuggingFace — Qwen3 embedding, 0.6B (Z890 / RDNA4 / 4090):

huggingface-cli download Qwen/Qwen3-Embedding-0.6B --local-dir ~/.cache/huggingface/qwen3-embed-0.6b

HuggingFace — Qwen3 embedding, 4B (Spark / 5090):

huggingface-cli download Qwen/Qwen3-Embedding-4B --local-dir ~/.cache/huggingface/qwen3-embed-4b

Ollama — Gemma 4 E4B (text + image input; GGUF via Ollama):

ollama pull gemma4:e4b

The Ollama tag gemma4:e4b serves text + image input, text output only — its model card lists Text, Image input. For audio-input Gemma 4 usage, serve the full weights via HF/transformers or vLLM (google/gemma-4-E4B-it), not the Ollama tag.

Ollama (8B on local node):

ollama pull hermes3:8b

Ollama (70B, Spark):

ssh pmoves-spark "ollama pull hermes3:70b"

HuggingFace — Unsloth Hermes 8B GGUF:

huggingface-cli download unsloth/Hermes-3-Llama-3.1-8B --local-dir ~/.cache/huggingface/hermes-v4-8b

HuggingFace — NeMo Omni VL (Spark, requires NeMo >= 2.1; ⚠ verify license first):

huggingface-cli download nvidia/NVLM-D-72B --local-dir ~/.cache/huggingface/nemo-omni-vl

Node routing rules

• dgx-spark (pmoves-spark): All models, primary for 70B+, Gemma 4 31B, and NeMo Omni
• 5090 / 4090: 8B models via Ollama (Unsloth GGUF preferred); Gemma 4 E4B (text+image via Ollama); Qwen3-Embedding-4B on 5090
• rdna4 (pmoves-rdna4): 8B via Ollama + ROCm; Qwen3-Embedding-0.6B via HF sentence-transformers
• z890: Qwen3-Embedding-0.6B only (no GPU large-model inference)

Verification

After pulling Hermes V4 8B:

ollama run hermes3:8b "respond: ready" --nowordwrap

After pulling a Qwen3 embedding (sentence-transformers loads directly):

python -c "from sentence_transformers import SentenceTransformer; \
m=SentenceTransformer('Qwen/Qwen3-Embedding-0.6B'); print(m.encode(['ready']).shape)"

Check TensorZero variant weight (staging, 0.0 until soak passes):

curl http://localhost:3030/api/v1/variants | jq '.[] | select(.name | contains("hermes_v4"))'