🌸 NanoSakura-2.2-0.2B

Japanese → Chinese Translation, Hybrid Transformer–Mamba2 Seq2Seq

▶️ Live Demo Space — try it in your browser, no setup required.

---

Overview

NanoSakura-2.2-0.2B is a highly efficient, custom-built Hybrid Sequence-to-Sequence (Seq2Seq) model for Japanese → Chinese translation. It pairs the deep contextual understanding of a Transformer Encoder with the fast, memory-efficient generation of a Mamba2 (State Space Model) Decoder.

With only ~188M parameters, it reaches translation quality competitive with — and on in-domain ACG text, well beyond — general-purpose LLMs nearly 10× its size, while keeping a hardware footprint small enough for edge deployment or high-throughput API serving.

💡 Which checkpoint should I use? This is the recommended, best-balanced checkpoint in the NanoSakura-2.x line. A later checkpoint (2.3) trades in-domain (ACG) quality for a small bump on general-domain benchmarks — see Evaluation for details.

Highlights

• 🏗️ Hybrid architecture — Transformer encoder + Mamba2 decoder with cross-attention, giving constant-memory per-step generation ($O(1)$ state update per token) instead of the usual growing KV-cache.
• 🎯 Domain specialist — trained on ACG (anime/comic/game)-domain distillation data; substantially outperforms general LLMs on in-domain translation (see shard_00134 results below).
• ⚖️ Balanced generalist — unlike later checkpoints, retains strong general-domain (FLORES-200) performance without sacrificing in-domain quality.
• 🪶 Tiny footprint — ~188M parameters, no "thinking"/CoT overhead required to hit its scores, unlike some LLM baselines.

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 with a Hybrid Architecture:

1. Encoder (Transformer): Self-Attention + RoPE + SwiGLU, fully capturing the global context of the source Japanese text in parallel.
2. Decoder (Mamba2 + Cross-Attention): Replaces decoder self-attention with Mamba2's State Space Model (SSM), enabling constant-memory generation — each decoding step is an $O(1)$ state update with no growing KV cache — while retaining Cross-Attention to accurately "look back" at the encoder's features and prevent hallucination.

Developed by	telecomadm1145
Model type	Hybrid Transformer–Mamba2 Seq2Seq
Language(s)	Japanese (ja) → Chinese (zh)
License	MIT
Parameters	~188M
Base model	NanoSakura-2-0.2B

🚀 Try it instantly

The fastest way to try the model is the hosted Space — no installation needed:

👉 huggingface.co/spaces/telecomadm1145/mamba2_trans

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 handles the $O(1)$ Mamba2 cache generation automatically.

ℹ️ Note: This model expects the source text to be manually terminated with an explicit <eos> token before encoding, as shown below.

import torch
from transformers import AutoModelForSeq2SeqLM, PreTrainedTokenizerFast

repo_id = "telecomadm1145/NanoSakura-2.2-0.2B"
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,
    torch_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 on the FLORES-200 benchmark (ja-zh) for general-domain quality, and on an in-domain ACG testset (shard_00134) for domain-specific quality. We report both lexical-overlap (SacreBLEU) and semantic-similarity (COMET) metrics. All results use greedy decoding.

⚠️ Metric details — please read before comparing:

• BLEU: All BLEU scores are computed with SacreBLEU using tokenize=zh (character-level tokenization), not spBLEU. Scores are internally consistent across every model and dataset in this table (all baselines were re-run under the same setup), but are not directly comparable to spBLEU numbers reported elsewhere (e.g., on the FLORES-200 leaderboard). Under zh tokenization, BLEU values are systematically higher than spBLEU.
• COMET: computed with Unbabel/wmt22-comet-da.

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	Qwen3-0.6B (thinking)	Qwen3-1.7B	Qwen3-1.7B (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

Reading the table:

• On in-domain ACG translation (shard_00134), NanoSakura-2.2 (0.8608 COMET) massively outperforms even Qwen3-1.7B-thinking (0.8182 COMET) — a model ~8.5× larger — despite the latter's extra chain-of-thought overhead. (The high absolute BLEU here reflects the combination of the zh tokenizer and the model's strong in-domain fit; compare across columns, not against external spBLEU numbers.)
• On general-domain translation (FLORES-200), 2.2 stays competitive with models several times its size, trailing Qwen3-1.7B by only ~0.007 COMET.

Recommended Use Cases

• ✅ ACG (anime/manga/game) text translation — visual novels, subtitles, in-game text, fan translation pipelines
• ✅ General-purpose ja→zh translation where a small, fast, edge-deployable model is preferred
• ✅ High-throughput API serving where Mamba2's constant-memory generation matters
• ⚠️ Not recommended for domains far outside ACG/general web text without further fine-tuning

License

Released under the MIT License.