cstr/harrier-270m-GGUF

harrier-270m GGUF

GGUF format of microsoft/harrier-oss-v1-270m for use with CrispEmbed and Ollama.

Files

File	Quantization	Size
harrier-270m-q4_k.gguf	Q4_K	0 MB
harrier-270m-q8_0.gguf	Q8_0	0 MB
harrier-270m.gguf	F32	0 MB

Recommended: Q8_0 for quality (cos vs HF: L2=1.0), Q4_K for size (L2=1.0).

Quick Start

CrispEmbed

./crispembed -m harrier-270m "Hello world"
./crispembed-server -m harrier-270m --port 8080

Ollama (with CrispStrobe fork)

echo "FROM harrier-270m-q8_0.gguf" > Modelfile
ollama create harrier-270m -f Modelfile
curl http://localhost:11434/api/embed -d '{"model":"harrier-270m","input":["Hello world"]}'

Python (CrispEmbed)

from crispembed import CrispEmbed
model = CrispEmbed("harrier-270m-q8_0.gguf")
vectors = model.encode(["Hello world", "Goodbye world"])

Model Details

Property	Value
Architecture	Gemma3
Parameters	270M
Embedding Dimension	640
Layers	18
Pooling	last-token
Tokenizer	SentencePiece BPE
Language	multilingual
Q8_0 vs HuggingFace	L2=1.0
Q4_K vs HuggingFace	L2=1.0

Server API

CrispEmbed server supports four API dialects:

• POST /embed -- native
• POST /v1/embeddings -- OpenAI-compatible
• POST /api/embed -- Ollama-compatible
• POST /api/embeddings -- Ollama legacy

Credits

• Original model: microsoft/harrier-oss-v1-270m
• Inference: CrispEmbed (MIT, ggml-based)