cheon-embedding-0.6b-v1

Highlights

• Top-3 class on the global multilingual benchmark at 0.6B. On the same 128 MMTEB (Multilingual, v2) tasks it averages 71.73. Only two much larger models score higher (27B and 12B). It is ahead of Qwen3-Embedding-8B, a model about 13 times its size, and of Gemini Embedding.
• Best among models of 8B or fewer parameters on these tasks: +1.76 over its 0.6B base model and +6.49 over Qwen3-Embedding-0.6B.
• Strong where multilingual search needs it: it beats the 8B model on bitext mining, classification, multilabel classification and retrieval.
• Long inputs and instructions: up to 32,768 tokens per text, with a one-line task instruction for queries.
• Drop-in loading: AutoModel with trust_remote_code=True, float32 weights, 1024-dimensional unit vectors (cosine similarity = dot product).

Global benchmark: MMTEB (Multilingual, v2)

Average main score over the same 128 of the 131 tasks. This model was scored with the official mteb harness (float32, up to 32,768 tokens); the other scores are from the official MTEB results repository. The remaining three tasks are being scored.

Ranks are among all models in the results repository that report all 128 tasks.

By task type (same 128 tasks):

Usage

import torch
from transformers import AutoModel, AutoTokenizer

repo = "cheonai/cheon-embedding-0.6b-v1"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModel.from_pretrained(repo, trust_remote_code=True, torch_dtype=torch.float32)
model.eval()

queries = model.encode(
    ["How do I renew my passport online?"],
    tok,
    instruction="Given a web search query, retrieve relevant passages that answer the query",
)
documents = model.encode(["Passports can be renewed online through the government portal."], tok)
scores = queries @ documents.T  # cosine similarity: the vectors are unit length

model(**tok(texts, padding=True, truncation=True, return_tensors="pt")).pooler_output gives the same vectors for texts that already carry their prefix.

Instructions

Queries take a one-line task description; documents take no prefix.

Instruct: {task description}
Query:{query text}

There is no space after Query:. encode(..., instruction=...) builds this prefix. Examples of task descriptions:

Specifications

Requirements

transformers>=4.51 and torch.

License

CC-BY-NC-4.0 — research and evaluation only. Contact us for commercial use. The base model (microsoft/harrier-oss-v1-0.6b) is MIT, but this fine-tune (learned weights) is ours and ships under the license above.