F2LLM-v2-0.6B — FP32 ONNX (dynamo export)

ONNX export of codefuse-ai/F2LLM-v2-0.6B, a general-purpose multilingual embedding model from the F2LLM-v2 family, trained on 60M high-quality multilingual examples supporting 200+ languages.

This is the full-precision (FP32) reference export. For production use, prefer the INT8 or INT4 variants which are 2–3× smaller with negligible quality loss.

Export method

Exported with torch.onnx.export(dynamo=True) (PyTorch 2.9, opset 20).

The dynamo exporter traces at the FX-graph / symbolic level. All internal tensor shapes — including the Qwen3 causal attention mask — carry symbolic batch and sequence dimensions throughout. Dynamic batch verified: batch = 1, 2, 4, 8 all produce correct output shapes.

Model details

Property	Value
Base model	codefuse-ai/F2LLM-v2-0.6B
Architecture	Qwen3 decoder
Parameters	~600 M
Embedding dim	1024
Max context	32 768 tokens
Languages	200+ (multilingual)
Inputs	`input_ids [batch, seq]`, `attention_mask [batch, seq]`
Output	`last_hidden_state [batch, seq, 1024]`
Pooling	Last-token pooling + L2 normalisation (applied by inference runtime)
File size	~2.4 GB (`model.onnx` + `model.onnx.data`)

Inference

import onnxruntime as ort
import numpy as np
from tokenizers import Tokenizer

tokenizer = Tokenizer.from_file("tokenizer.json")
tokenizer.enable_padding(pad_id=0, direction="right")
tokenizer.enable_truncation(max_length=512)

session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])

texts = ["semantic search example", "another sentence"]
enc  = tokenizer.encode_batch(texts)
ids  = np.array([e.ids            for e in enc], dtype=np.int64)
mask = np.array([e.attention_mask for e in enc], dtype=np.int64)

lhs = session.run(None, {"input_ids": ids, "attention_mask": mask})[0]  # [batch, seq, 1024]

# Last-token pooling: take embedding at last non-padding position
seq_lens   = mask.sum(axis=1) - 1
embeddings = lhs[np.arange(len(texts)), seq_lens]

# L2 normalise
norms      = np.linalg.norm(embeddings, axis=1, keepdims=True)
embeddings = embeddings / np.maximum(norms, 1e-8)
print(embeddings.shape)  # (2, 1024)

Query prefix: For asymmetric retrieval, prepend "Instruct: <task description>\nQuery: " to query strings. Documents are encoded without a prefix.

Files

File	Size	Description
`model.onnx`	~4 MB	ONNX graph (opset 20, no weights)
`model.onnx.data`	~2.38 GB	External weight data
`tokenizer.json`	8 MB	HuggingFace fast tokenizer
`config.json`	—	Model config

Quantized variants

Repo	Precision	Size	Notes
cstr/F2LLM-v2-0.6B-ONNX	FP32	2.4 GB	This repo — reference
cstr/F2LLM-v2-0.6B-ONNX-INT8	INT8 per-channel	1.1 GB	Recommended for most use
cstr/F2LLM-v2-0.6B-ONNX-INT4	INT4 MatMulNBits	0.9 GB	Minimum RAM
cstr/F2LLM-v2-0.6B-ONNX-INT8-FULL	INT8 incl. embeddings	0.6 GB	Smallest file

Citation

@misc{f2llm-v2,
      title={F2LLM-v2: Inclusive, Performant, and Efficient Embeddings for a Multilingual World},
      author={Ziyin Zhang and Zihan Liao and Hang Yu and Peng Di and Rui Wang},
      year={2026},
      eprint={2603.19223},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.19223},
}