Octen-Embedding-8B
by Octen
An 8B text embedding model on Qwen3, with a 40k token window
Octen/Octen-Embedding-8BDeploy Octen-Embedding-8B
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
A sentence-transformers embedding model built on the Qwen3 architecture, emitting 4096-dimensional vectors with a 40,960-token position limit. The long window is the reason to reach for it: most retrieval embedders cap at 512 or 8k tokens, so a contract, a deposition or a long transcript has to be chunked before it is embedded. This one can take considerably more of the document in a single pass, at the cost of an 8-billion-parameter forward pass per embedding.
Architecture
Qwen3 decoder, 36 hidden layers, 4096-dimensional hidden state, 151,665-token vocabulary, 40,960 max position embeddings. 7,567,295,488 parameters in safetensors. Pooling is configured in 1_Pooling/config.json rather than fixed by the architecture.
Mixpeek SDK Integration
# Bring your own vectors. Mixpeek does not run this model, so embed with it
# yourself and upsert the vectors beside the objects they describe.
from sentence_transformers import SentenceTransformer
from mixpeek import Mixpeek
model = SentenceTransformer("Octen/Octen-Embedding-8B")
mx = Mixpeek(api_key="API_KEY")
vector = model.encode("the full text of a long document") # 4096-d
mx.documents.upsert(
collection_id="col_your_collection",
documents=[{
"document_id": "filing-2026-0417",
"payload": {"source_key": "filings/2026/0417.pdf"},
"vectors": {"octen-8b": vector.tolist()},
}],
)Capabilities
- 4096-dimensional dense text embeddings
- 40,960-token position limit, so long documents need less chunking
- Runs through sentence-transformers, so it drops into existing embedding code
- Apache-2.0, which makes it usable in a commercial pipeline
Use Cases on Mixpeek
Frequently Asked Questions
Does Mixpeek run Octen-Embedding-8B for me?
No. Mixpeek runs its own extractors, and this is not one of them. You embed with it yourself and upsert the vectors, which is the bring-your-own-vectors path. The managed text extractor runs multilingual-e5-large-instruct at 1024 dimensions.
Is the 40k token window worth 8 billion parameters?
It depends entirely on your documents. If your passages fit comfortably in 512 tokens, a small model will be faster, cheaper and about as good. The long window earns its cost when chunking is destroying context that spans sections, which is common in contracts and filings and rare in product descriptions.
Can I use it commercially?
Yes. It is Apache-2.0 as of the 2026-09-16 sync. Check the model card before you deploy, because licences do change.
Specification
Build a pipeline with Octen-Embedding-8B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free