📄️ FFM Models Characteristics and Specifications
Version: 20250425
📄️ Llama3.3-FFM-70B-32K
Llama3.3-FFM-70B-32K is a large language model based on Meta Llama 3.3-70B with enhanced performance in Tool call. The accuracy of tool call has been significantly improved, with BFCL increasing from 85.08% of the base model to 92.18%. Both performance of Simple and Parallel saw substantial improvements.
📄️ Llama3.2-FFM-11B-V-32K
Llama3.2-FFM-11B-V-32K is a vision language model optimized for Traditional Chinese based on Meta-Llama-3.2-11B-Vision. It improves the shortcomings in handling Traditional Chinese from native model, making it the only multimodal model currently enhanced for Meta-Llama-3.2 in Traditional Chinese.
📄️ Llama3.1-FFM - 8B, 70B - 32K
Llama3.1-FFM is a large language model based on Meta Llama 3.1 with good performance in Traditional Chinese. By using the Llama3.1-FFM models, you agree to comply with the Meta-Llama3.1 licensing terms.
📄️ LLama3-FFM 70B、8B
Llama3-FFM series is an enhanced Traditional Chinese version of large language models based on the open-source large language model Meta Llama 3 launched by Meta. By using the Llama3-FFM series models, you agree to comply with the Meta-Llama3 licensing terms.
📄️ FFM-Mistral-7B、FFM-Mistral-7B-4K、FFM-Mixtral-8x7B
The FFM-Mistral series is an enhanced Traditional Chinese version of large language models based on the open-source large language models Mistral 7B and Mixtral 8x7B launched by Mistral AI. By using the FFM-Mistral series models, you agree to comply with the terms of the Apache 2.0 license.
📄️ FFM-Llama2-v2-7B, 13B, 70B (incl. FFM-Llama2)
FFM-Llama2 series models are enhanced Traditional Chinese versions of large language models based on the open-source large language model Meta Llama 2 launched by Meta.
📄️ FFM-embedding-v2(incl. v2.1), FFM-embedding
Embedding models can transform complex text into a set of vectors, converting it into a more manageable and understandable form while retaining key information. This helps in tasks such as text analysis, keyword analysis, and simple text content classification.
📄️ TAIDE-LX-7B
TAIDE-LX-7B is a local LLM created by Taiwan National Science and Technology Council. It selects local language corpora to understand local cultural semantics, enhancing the model's ability to process Traditional Chinese. It also incorporates texts from different domains to improve performance across various fields. Please agree their license before using. For more information, please refer to the official website.
📄️ Meta-Llama3.3-70B-Instruct-32K
Meta-Llama3.3-70B-Instruct is an open-source large language model released by Meta, whose performance is equivalent to Llama3.1-405B. Meta-Llama3.3-70B-Instruct is more cost-effective and significantly reduces computing requirements and usage thresholds.
📄️ Meta-Llama3.1 - 8B, 70B - 32K
Meta-Llama3.1 are open-source large language models released by Meta. Compared with Llama3, Llama3.1 has enhanced in terms of stronger reasoning capabilities, state-of-the-art tool use and multilingual translation. Meta-Llama3.1 models are provided with 32K context length in AFS. By using the Meta-Llama3.1 series models, you agree to comply with the Meta-Llama3.1 licensing terms.
📄️ Meta-Llama2-7B, 13B, 70B
Meta-Llama2 model is an open source, commercially available large language model developed and released by Meta. Meta-Llama2 has three sizes of parameters, 70 billion, 13 billion, and 7 billion, and can be used to complete tasks with different complexity. By using Meta-Llama2 series models, you agree to comply with the Meta-Llama2 licensing terms.
📄️ Meta-CodeLlama-7B, 13B, 34B-8K
Meta presented Code Llama, a family of LLMs for code generation and infilling derived from Meta-Llama2 and released under the same custom permissive license. CodeLlama provides state-of-the-art performance among open models, infilling capabilities, support for large input contexts, and zero-shot instruction following ability for programming tasks.
📄️ Jina-embeddings-v2-base-code
Jina-embeddings-v2-base-code is a multilingual embedding model that speaks English and 30 widely used programming languages. It supports 8192 sequence length and the model size is 161 million parameters. By using Jina-embeddings-v2-base-code, you agree to comply with the licensing terms. For more information, please refer to Hugging Face.
📄️ Qwen2.5-coder-7b-32k
Qwen2.5-Coder are large-scale language models specialized for code generation. The models support 92 programming languages and can handle various tasks related to coding, such as code generation, code reasoning and code fixing. Qwen2.5-coder models are not only enhancing coding capabilities but also maintaining its strengths in mathematics.
📄️ BAAI-bge-reranker-v2-m3
BAAI-bge-reranker-v2-m3 is a lightweight reranking model based on bge-m3, with 568 million parameters. It has multilingual capabilities, is easy to deploy, and offers fast inference speed.By using BAAI-bge-reranker-v2-m3, you agree to comply with the licensing terms.