FFM Models Characteristics and Specifications

Version: 20250425

Vision Language models	Model size	Context length	Multi-turn dialogue	OCRBench	MMStar	AI2D
Llama3.2-FFM	11B	32K	O	Surpass Meta-Llama3.2-11B	Surpass Meta-Llama3.2-11B	Equivalent to Meta-Llama3.2-11B

Chat models	Model size	Context length⁕	Expanded vocabulary list	Tool call	TMMLU+
Llama3.3-FFM	70B	32K	Equivalent to native model	Enhance Tool Call	Surpasses GPT4.0
Llama3.1-FFM	70B/8B	32K	Equivalent to native model	Enhance Tool Call	70B Surpasses GPT4.0
Llama3-FFM	70B/8B	8K	Equivalent to native model	Enhance Tool Call	70B Surpasses GPT4.0
FFM-Mixtral	MoE 8x7B	32K	Expand CN vocabulary list	Support	Surpasses GPT3.5
FFM-Mistral	7B	4K, 32K	Expand CN vocabulary list	Support	Surpasses GPT3.5
FFM-Llama2-v2	70B/13B/7B	4K	Expand CN vocabulary list	Support	70B Surpasses GPT3.5
FFM-Llama2	70B/13B/7B	4K	NA	NA	70B Surpasses GPT3.5

^⁕Context Length is the total number of tokens in the input plus the output.
^⁕⁕FFM model series all support multi-turn dialogue.

Embedding models	Model size	Context window	Embedding dimension	Language support	DRCD
FFM-embedding-v2.1	1B	8K	2048	zh-TW, English, German, French, Italian, Spanish, Portuguese, Hindi, Thai	Surpasses OpenAI embedding
FFM-Embedding-v2	1B	8K	2048	zh-TW, English, German, French, Italian, Spanish, Portuguese, Hindi, Thai	Surpasses OpenAI embedding
FFM-Embedding	1B	2K	1536	zh-TW, English	-