• Home
  • Categories
  • WIKI
  • LANGUAGE: en
    • [Seek Help] setup Spark x2.5 as a local ai
    UOS System 166 views · 2 replies ·
    Tofloor
    poster avatar
    OrangeJuice
    deepin
    2026-09-29 01:15
    Author

    I would like to use spark x2.5 as local ai in deepin, can someone tell me how to do it? Is it possible? What kind of file I need to download and how can I let deepin use it. Thx

    Reply Favorite View the author
    All Replies
    avatar
    deepin小助手
    Super Moderator
    OM
    2026-10-15 10:05
    #1

    thank you for your feedback

    Option 1: Ollama (recommended, command line)

    1. Install Ollama — open a terminal in deepin and run:

    curl -fsSL https://ollama.com/install.sh | sh

    This install method has been verified to work on deepin V23 . Ollama officially supports Debian-based systems .

    2. Make sure Ollama is v0.34.1 or newer — Spark-X2.5 uses a custom spark2_5 architecture, and official support was only added on 2026-09-15:

    ollama --version

    3. Run the model (no manual file download needed — Ollama pulls the weights automatically):

    # Lighter version (~1 GB, runs on ~8 GB RAM)
ollama run SparkLLM/Spark-X2.5-1.7B

# Or the stronger version (~2.5 GB quantized, ~16 GB RAM recommended)
ollama run SparkLLM/Spark-X2.5-4B

    That's it — you get a chat prompt right in the terminal. The 1.7B model officially runs on CPU-only machines; a GPU (NVIDIA with CUDA, or Huawei Ascend) just makes it faster .

    Option 2: LM Studio (GUI, no terminal)

    If you prefer a graphical app: install LM Studio 0.4.0+ (AppImage from lmstudio.ai works on deepin), search "Spark-X2.5" in the in-app model catalog, download, and click Load to chat. LM Studio added Spark-X2.5 support on 2026-09-08 .

    What file do you actually need to download?

    • With Ollama or LM Studio: nothing manually. The model files download automatically.
    • Manual download (only if you want the files yourself): grab the GGUF version from Hugging Face (XHToken/Spark-X2.5-4B-GGUF) or, if you're in China and HF is slow, ModelScope is the officially recommended mirror . Example:
    pip install -U modelscope
modelscope download --model XHToken/Spark-X2.5-4B-GGUF --local_dir ./spark

    Then you can import the .gguf file into Ollama via a Modelfile with FROM /path/to/model.gguf.

    Quick reality check

    • The 4B model runs well on a 12 GB GPU, or on CPU with 16 GB RAM — just slower .
    • Don't bother with the custom llama.cpp/Ollama source builds described in older guides — that was only needed before official spark2_5 support landed in September 2026.
    • The model has a "thinking" mode; use --think=false in Ollama if you want faster, direct answers .
    Reply View the author
    avatar
    OrangeJuice
    deepin
    2026-10-15 13:56
    #2
    deepin小助手：

    thank you for your feedback

    Option 1: Ollama (recommended, command line)

    1. Install Ollama — open a terminal in deepin and run:

    curl -fsSL https://ollama.com/install.sh | sh

    This install method has been verified to work on deepin V23 . Ollama officially supports Debian-based systems .

    2. Make sure Ollama is v0.34.1 or newer — Spark-X2.5 uses a custom spark2_5 architecture, and official support was only added on 2026-09-15:

    ollama --version

    3. Run the model (no manual file download needed — Ollama pulls the weights automatically):

    # Lighter version (~1 GB, runs on ~8 GB RAM)
ollama run SparkLLM/Spark-X2.5-1.7B

# Or the stronger version (~2.5 GB quantized, ~16 GB RAM recommended)
ollama run SparkLLM/Spark-X2.5-4B

    That's it — you get a chat prompt right in the terminal. The 1.7B model officially runs on CPU-only machines; a GPU (NVIDIA with CUDA, or Huawei Ascend) just makes it faster .

    Option 2: LM Studio (GUI, no terminal)

    If you prefer a graphical app: install LM Studio 0.4.0+ (AppImage from lmstudio.ai works on deepin), search "Spark-X2.5" in the in-app model catalog, download, and click Load to chat. LM Studio added Spark-X2.5 support on 2026-09-08 .

    What file do you actually need to download?

    • With Ollama or LM Studio: nothing manually. The model files download automatically.
    • Manual download (only if you want the files yourself): grab the GGUF version from Hugging Face (XHToken/Spark-X2.5-4B-GGUF) or, if you're in China and HF is slow, ModelScope is the officially recommended mirror . Example:
    pip install -U modelscope
modelscope download --model XHToken/Spark-X2.5-4B-GGUF --local_dir ./spark

    Then you can import the .gguf file into Ollama via a Modelfile with FROM /path/to/model.gguf.

    Quick reality check

    • The 4B model runs well on a 12 GB GPU, or on CPU with 16 GB RAM — just slower .
    • Don't bother with the custom llama.cpp/Ollama source builds described in older guides — that was only needed before official spark2_5 support landed in September 2026.
    • The model has a "thinking" mode; use --think=false in Ollama if you want faster, direct answers .

    Thank you for the time spent to answer, very well done. Thx appreciate!! 🙏

    Reply View the author