Local AI on Phones: Spark-X2.5 on OnePlus 8 Pro and Galaxy S26
Running AI locally on a phone provides privacy, offline access, predictable availability and freedom from cloud inference fees. Modern phones already contain powerful CPUs, GPUs and enough memory to become useful personal AI computers. The model and conversations remain on the device, while a local web server makes the AI accessible through an ordinary browser.
Spark-X2.5 is particularly interesting because its compact 4B size delivers capabilities normally associated with larger models: reasoning, coding, writing, translation, tool use and agentic workflows. It supports more than 200 languages and has a native context window of up to one million tokens, although usable phone context remains limited by memory, speed and battery consumption.
Its published scores are unusually strong for a model of this size. Spark-X2.5-4B scored 90.7 on AIME 2026 for mathematical reasoning and 65.1 on BFCL-V4 for tool calling. It also scored 54.6 on MCP-Atlas and 14.2 on MCP-Mark. MCP performance is important because it measures the model’s ability to work with external tools and data rather than functioning only as a chatbot.
We ran the approximately 2.6 GB Q4_K_M version entirely inside Termux, a Linux-style command-line environment for Android. The developer’s XHToken llama.cpp fork was cloned and compiled directly on both phones. This special fork was necessary because standard llama.cpp did not yet recognize the model’s spark2_5 architecture. Vulkan support was compiled into llama.cpp to use the phones’ Adreno GPUs.
On the OnePlus 8 Pro, CPU inference worked at approximately 12 prompt tokens/s and 4.6 generation tokens/s. Both the command-line interface and local web server worked, and the model successfully solved a second-order differential equation. Vulkan repeatedly crashed, even with only one layer offloaded. The likely limitation was the phone’s Adreno 650 and its old April 2022 Qualcomm Vulkan driver.
On the Galaxy S26, the same model ran successfully on both CPU and the Adreno 840 GPU. CPU performance reached approximately 37 prompt and 14 generation tokens/s. Full Vulkan offloading reached about 36 prompt and 20 generation tokens/s in the short test. Running through the web server with a 16K context produced approximately 33 prompt and 15 generation tokens/s.
The S26 CPU was roughly three times faster than the OnePlus CPU, while Vulkan produced more than four times the OnePlus generation speed. The relatively small difference between partial and full GPU offloading suggests that Spark-X2.5’s hybrid attention architecture is not yet fully optimized for Vulkan.
The experiment demonstrates that a capable reasoning and tool-oriented AI model can run completely locally on Android. The older OnePlus 8 Pro can operate it slowly through CPU inference, while the Galaxy S26 delivers practical interactive performance through Vulkan—without cloud inference, an internet connection or Qualcomm NPU execution.
https://huggingface.co/XHToken/Spark-X2.5-4B — Official Spark-X2.5 model page with architecture, capabilities, benchmark scores and deployment information.
https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF — Quantized GGUF versions, including the approximately 2.6 GB Q4_K_M model used on both phones.
https://github.com/XHToken/llama.cpp — Developer’s compatible llama.cpp fork required to run the spark2_5 architecture.
https://artificialanalysis.ai/hardware-inference-stack/mobile-phones — Artificial Analysis benchmark comparing the intelligence, speed and memory requirements of small models running on phones.
now latest llama cpp loads model without custom fork!!
Good performance, though some may prefer an aggressive quant of a MoE model