Repolode
Submit
XHToken

XHToken/Spark-X2.5

Open-weight 4B and 1.7B general-purpose language models from XHToken, aimed at on-device use and agentic tool-calling workflows.

View on GitHubWebsite
Stars
814
Forks
110
Open issues
5
Last commit
yesterday

What is it?

Spark-X2.5 is an open model series released in two sizes, 4B and 1.7B parameters, intended for conversation, writing, translation, reasoning, coding, tool use and agentic workflows. The models use a hybrid attention design that pairs one full-attention layer with three sliding-window attention layers, and they natively support context windows of up to 1M tokens.

The project provides weights on Hugging Face, ModelScope, Ollama and several regional platforms, along with FP8 and INT8 quantized variants. Deployment examples cover vLLM, SGLang, llama.cpp, MLX and LM Studio, including setups on NVIDIA GPUs and Huawei Ascend NPUs. A chat template and a tool-call parser are included for use with agent harnesses.

The repository holds the official model information, benchmark tables comparing against Qwen3.5 and Gemma4 models, quickstart commands for each runtime, and fine-tuning guidance using LLaMA-Factory. The weights are released under Apache 2.0.

Key features

  • 01Hybrid attention architecture with native context windows up to 1M tokens.
  • 02Available in 4B and 1.7B sizes, with FP8 and INT8 quantized versions.
  • 03Includes a tool-call parser and chat template for agent use.
  • 04Runs on vLLM, SGLang, llama.cpp, MLX, Ollama and LM Studio.
  • 05Supports NVIDIA GPUs and Huawei Ascend NPUs, with fine-tuning via LLaMA-Factory.

Use cases

  • Running a small general-purpose model locally on a laptop or edge device.
  • Serving an OpenAI-compatible API for coding or agent tools.
  • Processing long documents with a 1M-token context window.
  • Fine-tuning a compact model for a specific domain with LLaMA-Factory.

Quick start

  1. 1

    Install Ollama and run the model

    ollama run SparkLLM
  2. 2

    Or download the model from Hugging Face (official collection)

Things to know

  • The README does not document the exact Ollama model tag beyond the SparkLLM namespace, so check the Ollama page before running.
  • Benchmarks are self-reported by the publisher, and several comparisons are marked with asterisks indicating figures taken from other sources.
  • The README does not state the model license, though the metadata lists Apache-2.0.

Topics

#llm#ai-agent#on-device#quantization#vllm#llama-cpp#ollama#tool-calling#agent#ai

FAQ

What is Spark-X2.5?
Open-weight 4B and 1.7B general-purpose language models from XHToken, aimed at on-device use and agentic tool-calling workflows. Spark-X2.5 is an open model series released in two sizes, 4B and 1.7B parameters, intended for conversation, writing, translation, reasoning, coding, tool use and agentic workflows. The models use a hybrid attention design that pairs one full-attention layer with three sliding-window attention layers, and they natively support context windows of up to 1M tokens.
Is Spark-X2.5 free to use? What license does it have?
Spark-X2.5 is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is Spark-X2.5 still maintained?
Actively maintained. The last commit to Spark-X2.5 was on Oct 19, 2026.
What can I use Spark-X2.5 for?
Running a small general-purpose model locally on a laptop or edge device. Serving an OpenAI-compatible API for coding or agent tools. Processing long documents with a 1M-token context window. Fine-tuning a compact model for a specific domain with LLaMA-Factory.
How do I install Spark-X2.5?
Install Ollama and run the model: ollama run SparkLLM Or download the model from Hugging Face (official collection)
Where can I get Spark-X2.5?
The source code and installation instructions are on GitHub: https://github.com/XHToken/Spark-X2.5

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects