playtranslate/qwen2.5-1.5b-instruct-mnn

Qwen 2.5 1.5B Instruct — MNN format (PlayTranslate redistribution)

This repo packages taobao-mnn/Qwen2.5-1.5B-Instruct-MNN as a single zip for direct download by PlayTranslate's in-app catalog.

The model itself is unmodified — same llm.mnn / llm.mnn.weight / llm_config.json / tokenizer.txt files as the upstream taobao-mnn distribution, plus a config.json tuned for PlayTranslate's translation workload:

{
    "backend_type": "cpu",
    "thread_num": 4,
    "precision": "low",
    "memory": "low",
    "sampler_type": "greedy",
    "use_template": false,
    "reuse_kv": true
}

Notable runtime knobs:

• sampler_type: greedy — deterministic max-likelihood decoding (translation, not creative generation, so no temperature/top-p).
• use_template: false — bypass MNN's apply_chat_template. The upstream model ships without a jinja chat_template, and MNN's fallback strips role markers, so PlayTranslate's JNI bridge hand-builds the <|im_start|>system…<|im_end|> envelope before feeding it to the engine.
• reuse_kv: true — generate_init doesn't wipe the KV between calls, so PlayTranslate can prefill the system prompt once and rewind the KV via eraseHistory(sys_pos, current) between translations. Measured ~1.45× faster than the equivalent llama.cpp GGUF path on a Snapdragon 8 Gen 2 over a 500-sentence ja→en run (513 ms vs 746 ms median).

PlayTranslate downloads this as a single qwen2.5-1.5b-instruct-mnn.zip, SHA-256 verifies, then extracts via OnDeviceLlmDownloader's ZipExtract commit strategy.

Files

File	Size	Purpose
qwen2.5-1.5b-instruct-mnn.zip	~830 MB	The five-file MNN model bundle (zip).

Zip contents:

llm.mnn              (~1.1 MB, MNN computation graph)
llm.mnn.weight       (~828 MB, int4 quantized weights)
llm_config.json      (~400 B, MNN-side architecture metadata)
config.json          (~250 B, runtime config — the values above)
tokenizer.txt        (~3 MB, sentencepiece vocab)

Licensing

• Model weights: Qwen 2.5 1.5B Instruct, Apache-2.0 © Alibaba Cloud.
• MNN format conversion: taken from taobao-mnn/Qwen2.5-1.5B-Instruct-MNN, Apache-2.0 © Alibaba.

No further training or fine-tuning was applied; this is purely a packaging redistribution to support PlayTranslate's catalog/downloader interface.