Qwen

Qwen2.5-VL-3B-Instruct

A model by Qwen you run on your own computer, offline: how it did in our Arabic test, the size of each of its files as Hugging Face lists it, the memory a conversation takes worked out from the file itself as llama.cpp reserves it, and whether it fits your card and RAM.

Released:

Compare two products

The graphics card in your computer

Press the button and your browser reads the name of your computer's graphics card, which we pick in the check if the Lab has it.

The card's name is never sent to our site: it stays in this browser alone until you press "Forget my device".

How the name is read

Can your PC run it?

In GB, as Windows shows it under Settings > System > About.

Everything in the conversation, your questions and the model's answers. A token is a piece of a word.

Going by this page's figures

Pick the graphics card in your computer or type its RAM, and we'll say which of this model's files fit.

    We haven't worked out this model's conversation memory: its layer layout isn't checked against llama.cpp yet, so these sums count the file and working memory only, and it needs more.

    How we work out whether a model fits

    Specifications

    Specifications and their sources
    The model
    Parameters 3.8 billion[1]
    Longest conversation 128,000 tokens[2]
    GGUF files to download
    Q4_K_M file 1.93 GB[2]
    Q8_0 file 3.29 GB[3]
    The Arabic test
    Arabic score 70 %[4] Run by CoreITen from its Q4_K_M file, with llama.cpp
    Understanding Arabic text 80 %[4]
    Following Arabic instructions 50 %[4]
    Does not invent 81 %[4]
    Made-up products it gave specs for (of 8) 2[4]
    Answers that slipped into another script (of 56) 2[4]

    Change log

    1. Page published.

    Sources

    1. [1] Qwen/Qwen2.5-VL-3B-Instruct on Hugging Face Hugging Face, retrieved
    2. [2] ggml-org/Qwen2.5-VL-3B-Instruct-GGUF: Qwen2.5-VL-3B-Instruct-Q4_K_M.gguf Hugging Face, retrieved
    3. [3] ggml-org/Qwen2.5-VL-3B-Instruct-GGUF: Qwen2.5-VL-3B-Instruct-Q8_0.gguf Hugging Face, retrieved
    4. [4] CoreITen's Arabic test of local AI models, version 1 CoreITen, retrieved

    How we choose sources and check the numbers

    Compare two products