---
title: "Spark-X2.5 Beats Larger Rivals on Coding Agents While Running on a Laptop"
subtitle: "A new open-source 4B and 1.7B model family claims leading agentic and coding scores on-device, with native 1M-token context and Apache 2.0 licensing."
url: "https://alphasignal.ai/news/spark-x2-5-beats-larger-rivals-on-coding-agents-while-running-on-a-laptop"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-09-20T08:13:26.000Z"
updated: "2026-10-01T00:22:36.164Z"
type: "Repo"
topics:
  - "Llms"
  - "Open Source"
subtopics:
  - "Small Models"
  - "Long Context"
upvotes: 224
source_url: "https://github.com/XHToken/Spark-X2.5"
comments: 0
truncated: true
---

# Spark-X2.5 Beats Larger Rivals on Coding Agents While Running on a Laptop

> A new open-source 4B and 1.7B model family claims leading agentic and coding scores on-device, with native 1M-token context and Apache 2.0 licensing.

[Spark-X2.5](https://github.com/XHToken/Spark-X2.5) is a new Apache-2.0-licensed, open-weight model family designed for local coding, tool use, reasoning, multilingual chat, and other agentic workloads. The release includes 4-billion-parameter and 1.7-billion-parameter checkpoints on Hugging Face, ModelScope, and Ollama.

Both models support more than 200 languages and claim native context windows of up to 1,048,576 tokens. The project positions the 4B model for 16GB laptops, giving developers a potential route to private, offline agents without per-token API charges. Maximum context length, throughput, and memory use will depend heavily on the runtime, cache implementation, and weight precision.

* **Models:** Spark-X2.5-4B and Spark-X2.5-1.7B
* **License:** Apache 2.0
* **Context:** Up to 1,048,576 tokens
* **Languages:** More than 200
* **Primary workloads:** Coding, tool calling, long-document processing, reasoning, translation, and chat

## One million tokens, with caveats

Spark-X2.5 controls long-context costs with a hybrid attention architecture comprising one full-attention layer and three sliding-window layers. Full attention allows tokens to interact across the sequence, while sliding-window attention restricts most computation to nearby tokens.

That combination reduces the compute and key-value cache required by an architecture that applies full attention at every layer. Memory use still rises with prompt length, particularly in the full-attention layers, so architectural support for one million tokens does not guarantee that every laptop can process the maximum window.

The 4B checkpoint alone can fit within 16GB at suitable precision, but weights are only part of the memory budget. The runtime, KV cache, active prompt, generated tokens, and temporary buffers also consume memory. Teams planning to use the full window should benchmark their intended quantization, hardware, and serving stack before selecting a deployment target.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [XHToken/Spark-X2.5 GitHub Repository](https://github.com/XHToken/Spark-X2.5)
- [Spark-X2.5 Collection on Hugging Face](https://huggingface.co/collections/XHToken/spark-x25)
- [Spark-MLX-LLM Runtime](https://github.com/XHToken/Spark-MLX-LLM)