Post
37
Spark-X2.5 Running Locally on Two Android Phones
I compiled Spark-X2.5-4B Q4_K_M directly in Termux on a OnePlus 8 Pro and Samsung Galaxy S26 using the specialized XHToken llama.cpp fork.
This 2.6 GB model is attractive for private, offline phone AI because it supports reasoning, coding, translation, tool use and agentic workflows. Published scores include 90.7 on AIME 2026, 65.1 on BFCL-V4, 54.6 on MCP-Atlas and 14.2 on MCP-Mark.
The OnePlus CPU achieved 12 prompt and 4.6 generation tokens/s. It even solved a second-order differential equation, although Vulkan crashed—likely because of its old Adreno 650 driver.
The S26 CPU reached 37 prompt and 14 generation tokens/s. Vulkan on its Adreno 840 reached about 36 prompt and 20 generation tokens/s. Its 16K-context web server delivered 33 prompt and 15 generation tokens/s.
A capable reasoning and tool-oriented AI can now run privately and offline on a phone, without cloud inference or NPU execution.
I compiled Spark-X2.5-4B Q4_K_M directly in Termux on a OnePlus 8 Pro and Samsung Galaxy S26 using the specialized XHToken llama.cpp fork.
This 2.6 GB model is attractive for private, offline phone AI because it supports reasoning, coding, translation, tool use and agentic workflows. Published scores include 90.7 on AIME 2026, 65.1 on BFCL-V4, 54.6 on MCP-Atlas and 14.2 on MCP-Mark.
The OnePlus CPU achieved 12 prompt and 4.6 generation tokens/s. It even solved a second-order differential equation, although Vulkan crashed—likely because of its old Adreno 650 driver.
The S26 CPU reached 37 prompt and 14 generation tokens/s. Vulkan on its Adreno 840 reached about 36 prompt and 20 generation tokens/s. Its 16K-context web server delivered 33 prompt and 15 generation tokens/s.
A capable reasoning and tool-oriented AI can now run privately and offline on a phone, without cloud inference or NPU execution.