MITRA-QWEN is a 9B translation model for classical Buddhist languages, built on Qwen/Qwen3.5-9B. It is the translation engine of the Dharmamitra project and is trained specifically to translate Classical Tibetan, Sanskrit, Pāli, and Classical/Buddhist Chinese into English, Japanese, Modern Chinese, Korean and several European languages.

It was produced by continued pretraining on a large domain corpus, instruction fine-tuning on a curated, quality-filtered parallel set, and a final preference-optimization (CPO) stage.

Under blind pairwise evaluation it matches or outperforms Claude Opus 4 on these language pairs (see Evaluation), while running as a self-hosted 9B model on a single GPU.

Supported directions

Source: Classical Tibetan, Sanskrit, Pāli, Classical/Buddhist Chinese (and their reverses). Target: English, Japanese, Modern Chinese, Korean, German, French, and others.

Prompt template

The model is trained on a specific instruction wrapped inside the Qwen chat template (as the user turn). Reasoning/thinking should be disabled (enable_thinking=False). There are two modes: closed-book, and in-context (ICL) where retrieved reference translations are supplied.

1. Closed-book (no in-context examples)

Put this text as the user message content:

You are an expert translator of classical Asian languages. In your translation, make sure to use proper IAST diacritics if Sanskrit terms occur, but also translate Sanskrit terms into English if English is the target language. The translation should be fluid and accurate. If the input is in English and the target is English, just return the input then.

Here is a piece of text to translate: {SOURCE_TEXT}
Provide only the translation, without any explanation or additional information. Provide your translation in {TARGET_LANGUAGE}:

{TARGET_LANGUAGE} is the full name of the target, e.g. English, Japanese, Modern Chinese, Classical Chinese, Tibetan, Sanskrit.

2. In-context / retrieval-augmented (ICL)

If you have reference translations (e.g. from a fuzzy-match / kNN index), insert them as SOURCE Means in TARGET_LANGUAGE: TRANSLATION lines before the text to translate:

You are an expert translator of classical Asian languages. In your translation, make sure to use proper IAST diacritics if Sanskrit terms occur, but also translate Sanskrit terms into English if English is the target language. The translation should be fluid and accurate. Here is a section of reference translations. Make use of them to improve your translation, but only if they make sense and really help. Never reproduce them verbatim; always paraphrase them.
{REF_SOURCE_1} Means in {TARGET_LANGUAGE}: {REF_TRANSLATION_1}
{REF_SOURCE_2} Means in {TARGET_LANGUAGE}: {REF_TRANSLATION_2}
Here is a piece of text to translate: {SOURCE_TEXT}
Provide only the translation, without any explanation or additional information. Provide your translation in {TARGET_LANGUAGE}:

The reference block is optional. On short inputs it makes little difference; on longer / domain-specific passages it helps anchor terminology. The model is robust to minor wording changes in the instruction.

Usage

PyTorch / 🤗 Transformers

Requires transformers >= 5.13 (native Qwen3.5 support).

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "buddhist-nlp/mitra-qwen35-translate"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="cuda")

def translate(source, target_language="English", references=None):
    header = ("You are an expert translator of classical Asian languages. In your translation, "
              "make sure to use proper IAST diacritics if Sanskrit terms occur, but also translate "
              "Sanskrit terms into English if English is the target language. The translation "
              "should be fluid and accurate.")
    if references:  # list of (ref_source, ref_translation)
        header += (" Here is a section of reference translations. Make use of them to improve your "
                   "translation, but only if they make sense and really help. Never reproduce them "
                   "verbatim; always paraphrase them.")
        header += "\n" + "\n".join(f"{s} Means in {target_language}: {t}" for s, t in references)
    prompt = (f"{header}\nHere is a piece of text to translate: {source}\n"
              f"Provide only the translation, without any explanation or additional information. "
              f"Provide your translation in {target_language}:")
    text = tok.apply_chat_template([{"role": "user", "content": prompt}],
                                   tokenize=False, add_generation_prompt=True, enable_thinking=False)
    ids = tok(text, return_tensors="pt").to(model.device)
    out = model.generate(**ids, max_new_tokens=1024, do_sample=False)
    return tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True).strip()

# closed-book
print(translate("bcom ldan 'das kyi zhabs la phyag 'tshal lo", "English"))
# with in-context references
print(translate("buddho bhagavā", "English",
                references=[("buddho bhagavā arahaṃ", "the Buddha, the Blessed One, the worthy one")]))

vLLM

MITRA-QWEN uses the Qwen3.5 architecture, so serve it exactly as you would the base Qwen/Qwen3.5-9B:

vllm serve buddhist-nlp/mitra-qwen35-translate --max-model-len 32768

Then send the same instruction (from the template above) as the user message to the OpenAI-compatible /v1/chat/completions endpoint with chat_template_kwargs={"enable_thinking": false}. Requires a vLLM build with native Qwen3.5 hybrid-attention support. FP8 + FP8 KV-cache (--quantization fp8 --kv-cache-dtype fp8) gives ~1.5× throughput on Blackwell GPUs at no measurable quality loss.

Recommended decoding: temperature=0 (greedy) or temperature=0.2, with a small repetition_penalty (1.05) to suppress the rare long-transliteration repetition loop.

Evaluation

Blind pairwise A/B against Claude Opus 4 (raw, no retrieval), on held-out real-traffic-length passages, 100 comparisons per language pair (500 total). Two independent LLM judges; A/B order randomized per item. Numbers are MITRA-QWEN's win rate (share of comparisons the judge preferred MITRA-QWEN).

Language pair	Judge: gemini-3.1-flash-lite	Judge: claude-sonnet-4
Sanskrit → Japanese	81%	55%
Tibetan → Modern Chinese	65%	59%
Tibetan → English	62%	51%
Classical Chinese → English	60%	54%
Pali → English	52%	44%
Overall	64.0% (Elo +100)	52.6% (Elo +18)

MITRA-QWEN wins or ties on every language pair under both judges. Against the previous Gemini-3-flash-preview production pipeline it wins 57.8% overall (claude-sonnet-4 judge, Elo +55).

Evaluation uses gemini-3.1-flash-lite and claude-sonnet-4 as automated judges; scores reflect each judge's blind pairwise preference, not a fixed reference metric.

Translation into classical languages

Where MITRA-QWEN differs most from general-purpose LLMs is translation into the classical languages and registers of the Buddhist canon — Buddhist/Classical Chinese, Classical Tibetan and Sanskrit — which require canonical terminology and register that general models do not reproduce.

Scores below are GEMBA (LLM-as-judge translation quality, 0–100, reference-based, gemini-3.1-flash-lite) on held-out canonical test sets, 300 sentence pairs per direction:

Direction	MITRA-QWEN	gemini-3.6-flash	gemini-3-flash-preview	gemini-3.5-flash
Sanskrit → Buddhist Chinese	84.2	65.3	62.7	53.5
Tibetan → Buddhist Chinese	85.7	75.4	75.3	73.9
Tibetan → Modern Chinese	86.9	85.9	85.3	85.4

MITRA-QWEN leads by +10 to +19 GEMBA points on the canonical-Chinese directions.

Additional directions (MITRA-QWEN, same protocol):

Direction	GEMBA
Sanskrit → Tibetan	86.3
Tibetan → Sanskrit	81.4
Chinese → Tibetan	80.4
Chinese → Sanskrit	79.3
Sanskrit → English	85.1
Tibetan → English	84.6
Classical Chinese → English	84.4
Sanskrit → Japanese	83.4
Pali → English	82.2

Training

Base: Qwen/Qwen3.5-9B (hybrid linear-attention + full-attention, 262k context).
Continued pretraining: ~22B tokens of domain text (Tibetan, Sanskrit, Pāli, Buddhist Chinese corpora, parallel data, linguistic annotations), with English retained to prevent forgetting.
Instruction fine-tuning: a curated, quality-filtered parallel set — every training target judge-scored and, where a stronger candidate existed, replaced by a higher-quality translation; a mix of closed-book and retrieval-augmented prompts, plus document-level and post-editing tasks.
Preference optimization (CPO): preference triplets scored target-side by an ensemble judge, with a behavior-cloning anchor (no plain DPO).

Limitations

Best results require the prompt template above with thinking disabled.
Standard n-gram metrics (BLEU/chrF) understate quality here: the model uses IAST diacritics and a modern scholarly register that diverge from older reference conventions. Prefer semantic / LLM-judge / human evaluation.
Reverse directions into Sanskrit and very low-resource pairs are weaker.

Acknowledgement

This model is released to support the wider community of scholars, translators, and digital-humanities projects working with classical Buddhist languages. If MITRA-QWEN is helpful to your work — whether it powers part of your translation infrastructure, or contributes to published translations — we would be grateful if you would acknowledge it. A short credit to the Dharmamitra / MITRA project (with a link to https://dharmamitra.org) is very much appreciated and helps sustain this open, non-commercial effort. It is a kind request rather than a requirement, and it means a great deal to a small team.

Citation

If you use this model, please cite the Dharmamitra / MITRA project. See https://dharmamitra.org.