GeeVec-Embeddings-1.0-Lite

GeeVec-Embeddings-1.0-Lite is a lightweight domain-adaptive text embedding model, with only 0.35B activated parameters, built on top of a Qwen3-style base model using a PseudoMoE architecture. It is optimized for retrieval tasks and supports domain routing for improved specialization:

• general: the default route, suitable for general-purpose multilingual retrieval.
• coding: specialized for code-related retrieval, including programming concepts, APIs, and technical documentation.
• reasoning: specialized for tasks that require deeper semantic understanding, multi-step inference, and complex query matching.

Despite its compact size, GeeVec-Embeddings-1.0-Lite delivers strong performance across a wide range of benchmarks. It achieves SOTA performance among small-size (<1B) models on the MMTEB(Multilingual, v2) retrieval task, with an nDCG@10 score of 74.66 (as of 2026/04/02). It also performs competitively on MMTEB(eng, v2), BEIR, CoIR, and BRIGHT, demonstrating strong retrieval capability despite its lightweight design.

Meanwhile, we also provide an API service for a larger 8B-scale model, GeeVec-Embeddings-1.0. Like GeeVec-Embeddings-1.0-Lite, it is optimized for retrieval tasks and supports the same three domains: general, coding, and reasoning. GeeVec-Embeddings-1.0 achieves SOTA performance on the MMTEB(Multilingual, v2) retrieval task, with an nDCG@10 score of 81.18 (as of 2026/04/02). API usage documentation: https://www.geevec.com/documentation.

Introduction

This repository hosts the model geevec-embeddings-1.0-lite.

Technical highlights:

• Model Type: Text Embedding
• Total Parameters: 349M activated / 366M total
• Context Length: 32,768
• Embedding dimension: Up to 4096, supports user-defined output dimensions ranging from 256 to 4096 (recommended dimensions: 256, 512, 1024, 2048, 4096)
• Domain-specific support: general (default), coding, reasoning
• Pooling Method: last-token pooling

Usage

Using FlagEmbedding

git clone https://github.com/FlagOpen/FlagEmbedding.git
cd FlagEmbedding
pip install -e .

from FlagEmbedding import FlagAutoModel

model_path = "geevec-ai/geevec-embeddings-1.0-lite"

model = FlagAutoModel.from_finetuned(
    model_path,
    model_class="decoder-only-pseudo_moe",
    query_instruction_for_retrieval="Given a question, retrieve passages that answer the question.",
    query_instruction_format="Instruct: {}\nQuery: {}",
    domain_for_pseudo_moe="general",  # general / coding / reasoning
    use_bf16=True,
    use_fp16=False,
    trust_remote_code=True,
    devices="cuda:0",  # if you do not have a GPU, set this to "cpu"
)

queries = [
    "how much protein should a female eat",
    "summit define",
]
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day.",
    "Definition of summit for English Language Learners: the highest point of a mountain; the highest level; a meeting between leaders.",
]

query_embeddings = model.encode_queries(queries)
document_embeddings = model.encode_corpus(documents)

similarity = query_embeddings @ document_embeddings.T
print(similarity)

Using Sentence Transformers

from sentence_transformers import SentenceTransformer
import torch

model_path = "geevec-ai/geevec-embeddings-1.0-lite"

# Load with trust_remote_code=True because the model defines custom modules.
model = SentenceTransformer(
    model_path,
    model_kwargs={"torch_dtype": torch.bfloat16},
    trust_remote_code=True,
)

queries = [
    "How can I optimize a Python function that has nested loops?",
    "What is the difference between eigenvalue decomposition and SVD?",
]

documents = [
    "Use vectorization, caching, and algorithmic improvements to reduce complexity.",
    "Eigenvalue decomposition applies to square matrices; SVD works for any matrix.",
]

# Optional domain routing: general / coding / reasoning
query_embeddings = model.encode(queries, domain="coding", normalize_embeddings=True)
doc_embeddings = model.encode(documents, domain="coding", normalize_embeddings=True)

similarity = query_embeddings @ doc_embeddings.T
print(similarity)

Using HuggingFace Transformers

import torch
import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel


def last_token_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]


def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery: {query}'


task = 'Given a web search query, retrieve relevant passages that answer the query.'
queries = [
    get_detailed_instruct(task, "How can I optimize a Python function that has nested loops?"),
    get_detailed_instruct(task, 'summit define')
]
# No need to add instructions for documents
documents = [
    "Use vectorization, caching, and algorithmic improvements to reduce complexity.",
    "What is the difference between eigenvalue decomposition and SVD?",
]
input_texts = queries + documents

tokenizer = AutoTokenizer.from_pretrained("geevec-ai/geevec-embeddings-1.0-lite")
model = AutoModel.from_pretrained("/geevec-ai/geevec-embeddings-1.0-lite", trust_remote_code=True)
model.eval()

max_length = 4096
# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt', pad_to_multiple_of=8)

with torch.no_grad():
    outputs = model(**batch_dict)
    embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
    
# normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())

Notes

• This model uses custom files modeling_qwen3_pseudo_moe.py, configuration_qwen3_pseudo_moe.py, and pseudo_moe_st_module.py.
• When loading from local path or hub, set trust_remote_code=True.
• If you do not specify a domain, the model uses general by default.

Evaluation

The following benchmark results summarize the performance of GeeVec-Embeddings-1.0 and GeeVec-Embeddings-1.0-Lite on the main retrieval and embedding evaluation suites.

MMTEB(Multilingual, v2) - general

MMTEB(eng, v2) - general

BEIR - general

CoIR - coding

BRIGHT - reasoning