聯發創新基地(MediaTek Research) 致力於研究基礎模型。我們將研究體現在適合正體中文使用者的模型上,並在使用權許可的情況下,提供模型給學術界研究或產業界使用。
[2025.01.24] Breeze 2 系列,包含多模態影像語言模型 模型開源 及 論文發表; 以及語音合成模型 模型開源 及 論文發表
[2024.09.23] Breeze FC 模型開源 及 論文發表
[2024.08.10] BreeXe 模型開源
[2024.07.03] 發布 MR 模型使用套件 mtkresearch 並上架到 PyPi
[2024.05.24] Generative Fusion Decoding (GFD) 技術發表 以及 Breezper 實作程式碼發布
[2024.01.12] Breeze-7B 模型開源 及 論文發表
[2023.10.20] 開放繁體中文評測 TC-Eval (new version: TCEval-v2)
[2023.09.14] Model 7 - C 開放試用 及 論文
[2023.08.15] Model 7 - B 開放試用
[2023.04.10] 開源 Bloom-zh 3B 模型 及 論文
[2023.03.07] 開源 Bloom-zh 1B1 模型 及 論文
【Paper】【Kaggle Demo】【Collection】
【TTS Paper】【TTS Kaggle Demo】【TTS GitHub】
Breeze 2 是一套先進的多模態模型家族。包含多模態語言模型以及語音合成模型。多模態語言模型提供 3B 和 8B 參數配置,專為加強繁體中文語言表示而設計。在 LLaMA 3.2 的基礎上,Breeze 2 持續在大規模語料庫上進行預訓練,以進一步加強繁體中文的語言與文化內涵。該模型結合了視覺編碼器與橋接模組,實現了視覺感知能力,同時通過提示模板與函數調用數據的後訓練,支持函數調用功能。語音合成模型支援繁體中文、英文以及注音輸入,中文語音輸出,並支援聲音複製(Voice Cloning)。
English Content
The Breeze 2 Herd of Models: Traditional Chinese LLMs Based on LLaMA with Vision-Aware and Function-Calling CapabilitiesBreeze 2 is a suite of advanced multi-modal models, including multi-modal language models and speech synthesis models. The multi-modal language models are available in 3B and 8B parameter configurations, specifically designed to enhance Traditional Chinese language representation. Building upon the LLaMA 3.2, Breeze 2 continues pretraining on an extensive corpus to further enhance the linguistic and cultural heritage of Traditional Chinese. It incorporates vision-aware capabilities through a visual encoder and a bridge module, and supports function-calling via prompt templates and post-training on function-calling data. The speech synthesis model supports Traditional Chinese, English, and Bopomofo input, with Chinese speech output, and also supports voice cloning.
Breexe-8x7B 是一個語言模型家族,基於 Mixtral-8x7B 開發,專門針對繁體中文使用。
Breexe-8x7B-Base 是 Breexe-8x7B 系列的基礎模型。Breexe-8x7B-Base 擴展了原始詞彙表,新增了 30,000 個繁體中文詞彙。在詞彙表擴展的情況下,Breexe-8x7B 在繁體中文推理速度上是 Mixtral-8x7B 的兩倍。
Breexe-8x7B-Instruct 是基於 Breexe-8x7B-Base 的衍生模型,使得該模型可以直接用於常見任務,如問答、檔案檢索生成(RAG)、多輪對話和摘要。 Breexe-8x7B-Instruct 在繁體中文和英文基準測試中表現出色,與 OpenAI 的 gpt-3.5-turbo-1106 相媲美。
English Content
Breexe-8x7B is a language model family that builds on top of Mixtral-8x7B, specifically intended for Traditional Chinese use.Breexe-8x7B-Base is the base model for the Breexe-8x7B series. Breexe-8x7B-Base expands the original vocabulary with additional 30,000 Traditional Chinese tokens. With the expanded vocabulary, Breexe-8x7B operates at twice the inference speed for Traditional Chinese to Mixtral-8x7B.
Breexe-8x7B-Instruct derives from the base model Breexe-8x7B-Base, making the resulting model amenable to be used as-is for commonly seen tasks, such as Q&A, RAG, multi-round chat, and summarization. Breexe-8x7B-Instruct demonstrates impressive performance in benchmarks for Traditional Chinese and English, on par with OpenAI's gpt-3.5-turbo-1106.
【Paper】【Collection】
Breeze-7B 是一個語言模型家族,基於 Mistral-7B 開發,專門針對繁體中文使用。 有關此模型的詳細資訊,請參閱我們的論文。
實用性方面:
- Breeze-7B-Base 擴增了原始詞表,新增了 30,000 個繁體中文詞彙。在詞彙表擴增且其他條件相同的情況下,Breeze-7B 在繁體中文推理速度上是 Mistral-7B 和 Llama 7B 的兩倍。
- Breeze-7B-Instruct 可直接用於常見任務,如問答、檔案檢索生成 (RAG)、多輪對話和摘要。
性能方面:
- Breeze-7B-Instruct 在繁體中文和英文基準測試中表現出色,與同類型的開源模型如 Taiwan-LLM-7B/13B-chat、QWen(1.5)-7B-Chat 和 Yi-6B-Chat 相比,具有顯著優勢。
English Content
Breeze-7B is a language model family that builds on top of Mistral-7B, specifically intended for Traditional Chinese use.For details of this model please read our paper.
Practicality-wise:
- Breeze-7B-Base expands the original vocabulary with an additional 30,000 Traditional Chinese tokens. With the expanded vocabulary, and everything else being equal, Breeze-7B operates at twice the inference speed for Traditional Chinese to Mistral-7B and Llama 7B.
- Breeze-7B-Instruct can be used as is for common tasks such as Q&A, RAG, multi-round chat, and summarization.
Performance-wise:
- Breeze-7B-Instruct demonstrates impressive performance in benchmarks for Traditional Chinese and English when compared to similar-sized open-source contemporaries such as Taiwan-LLM-7B/13B-chat, QWen(1.5)-7B-Chat, and Yi-6B-Chat.