Jarvis Voice’s pinned EmbeddingGemma 300M BF16 artifact: 768-dimensional multilingual embeddings using one immutable model across cloud and local Ollama deployments.

Jarvis Embedding Ollama Artifact

Github repo

https://github.com/bigsk1/jarvis-voice

Jarvis uses one versioned Ollama embedding artifact in cloud and local modes:

bigsk1/jarvis-embedding:bf16-v1

Pull it on every daemon listed in OLLAMA_BASE_URL:

ollama pull bigsk1/jarvis-embedding:bf16-v1

Artifact identity

Field	Pinned value
Registry tag	`bigsk1/jarvis-embedding:bf16-v1`
Ollama manifest SHA-256	`85462619ee721b466c5927d109d4cb765861907d5417b9109caebc4e614679f1`
Model layer SHA-256	`0800cbac9c2064dde519420e75e512a83cb360de3ad5df176185dc69652fc515`
Upstream artifact	Ollama `embeddinggemma:300m-bf16`
Architecture	Gemma 3 embedding model
Parameters	307.58M
Quantization	BF16
Dimensions	768
Context window	2048 tokens
Pooling	Mean

The tag was published as an exact manifest copy of the upstream BF16 artifact, including its Gemma Terms of Use license layer, template, parameters, and model weights. Representative query, document, and similarity inputs produced bit-for-bit identical normalized vectors before publication.

Hugging Face recovery mirror

The exact GGUF model layer and Ollama provenance artifacts are also archived in the public, automatically gated Hugging Face repository:

https://huggingface.co/bigsk1/jarvis-embedding-GGUF

The mirror contains the 621,867,104-byte BF16 GGUF, the original Ollama manifest, config and parameter layers, the bundled Gemma terms, the required redistribution notice, a recovery Modelfile, and SHA-256 checksums. Hugging Face is a disaster-recovery source, not a second Jarvis runtime provider; Jarvis continues to call Ollama exclusively.

An Ollama 0.32.13 recovery test produced bit-for-bit identical 768-dimensional vectors for representative query, document, and similarity inputs. However, ollama create reserialized the GGUF and added a template layer, so the rebuilt Ollama manifest digest differed from the pinned digest. A rebuild from the Hugging Face Modelfile must therefore use a new immutable tag, update the Jarvis fingerprint, and re-embed every persisted namespace. It must not replace the existing bf16-v1 identity.

Immutability policy

Do not overwrite bf16-v1. A future weight, template, parameter, or license change must use a new versioned tag such as bf16-v2, update OLLAMA_EMBEDDING_MODEL_DIGEST, and rebuild every persisted embedding namespace. Jarvis verifies both the exact tag and manifest digest and fails closed rather than mixing artifacts.

There is intentionally no runtime dependency on the upstream embeddinggemma:latest tag and no Jarvis latest tag in the configuration.

Verification

After pulling the model, verify every configured host without opening or creating a database:

./bin/check-embeddings-health.py --both --runtime-only

To inspect one daemon directly:

curl -fsS http://localhost:11434/api/tags \
  | jq '.models[] | select(.name == "bigsk1/jarvis-embedding:bf16-v1") \
        | {name, digest, size, details}'

The reported digest must equal the pinned manifest SHA-256 above.