This model is a highly efficient, custom-built Hybrid Sequence-to-Sequence (Seq2Seq) model designed for Japanese to Chinese translation. It combines the deep understanding capabilities of a Transformer Encoder with the blazing-fast, memory-efficient generation of a Mamba2 (State Space Model) Decoder.

Model Details

Model Description

Traditional Seq2Seq models (like T5 or BART) rely entirely on Transformers. While powerful, the self-attention mechanism in the decoder leads to an $O(N^2)$ computational bottleneck and high KV-cache memory usage during text generation.

This model solves that by introducing a Hybrid Architecture:

Encoder (Transformer): Uses Self-Attention + RoPE + SwiGLU to fully capture the global context of the source Japanese text in parallel.
Decoder (Mamba2 + Cross-Attention): Replaces self-attention with Mamba2's State Space Model (SSM). This allows for $O(1)$ state-updating generation (no growing KV cache), while retaining Cross-Attention to accurately "look back" at the encoder's features to prevent hallucination.

With only ~188 Million parameters, it achieves excellent translation quality while maintaining an extremely low hardware footprint, making it ideal for edge deployment or high-throughput API services.

Developed by: telecomadm1145
Model type: Hybrid Transformer-Mamba2 Seq2Seq
Language(s) (NLP): Japanese (ja), Chinese (zh)
License: MIT
Parameters: ~188M

How to Get Started with the Model

Because this model uses a custom architecture, you must use trust_remote_code=True when loading it with the transformers library. The custom modeling_mamba2_s2s.py will handle the $O(1)$ Mamba2 cache generation automatically.

Use the code below to get started:

python

import torch
from transformers import AutoModelForSeq2SeqLM, PreTrainedTokenizerFast

repo_id = "telecomadm1145/NanoSakura-2.3-0.2B"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = PreTrainedTokenizerFast.from_pretrained(repo_id)
model = AutoModelForSeq2SeqLM.from_pretrained(
    repo_id, 
    trust_remote_code=True, 
    dtype=torch.float32 
)
model.to(device)

text = "おはようございます、今日の天気はいいですね！"
input_ids = tokenizer.encode(text + "<eos>")
input_tensor = torch.tensor([input_ids]).to(device)

output_ids = model.generate(
    input_tensor, 
    max_new_tokens=256,
)
result = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(f"Translation: {result}")
# Output: 早安，今天的天气真好呢！

Evaluation

We evaluated the model on the FLORES-200 benchmark (Japanese to Chinese, ja-zh) and a testset(shard_00134). To provide a comprehensive assessment, we report both lexical-overlap metrics (SacreBLEU) and semantic-similarity neural metrics (COMET).

Metrics

Metric	opus-mt-ja-zh(~73M)	NanoSakura-2-0.2B	NanoSakura-2.2-0.2B	NanoSakura-2.3-0.2B	NanoSakura-0.3B	nllb-200-1.3B	Qwen3-0.6B(fp16)	Qwen3-0.6B(fp16,thinking)	Qwen3-1.7B(fp16)	Qwen3-1.7B(fp16,thinking)
FLORES-200 BLEU	25.67	23.27	26.55	28.67	22.36	20.87	12.58	21.13	27.12	27.56
FLORES-200 COMET	0.8371	0.8380	0.8494	0.8563	0.8307	0.7805	0.8020	0.8220	0.8561	0.8571
shard_00134 BLEU	8.07	58.13	57.55	49.32	58.71	5.73	6.89	14.57	23.37	24.60
shard_00134 COMET	0.4493	0.8615	0.8608	0.8558	0.8654	0.5181	0.6930	0.7414	0.8158	0.8182

All evaluated using greedy decoding.