Spark-X2.5 on Unsloth and LM Studio: a real bilingual local demo

Description and use

Spark-X2.5 ran locally on both Unsloth and LM Studio, producing consistent bilingual answers to a math prompt.

Problem

Demonstrate that Spark-X2.5 can run locally in both Unsloth llama.cpp and LM Studio, producing consistent bilingual outputs for a math problem.

Outcome

Spark-X2.5 successfully ran locally on both Unsloth and LM Studio, producing consistent bilingual outputs for a math prompt, demonstrating local deployment feasibility.

Workflow

• 1. Download Spark-X2.5 GGUF checkpoint (e.g., Spark-X2.5-1.7B Q4_K_M).
• 2. Set up Unsloth llama.cpp backend (build b10840, 4 CPU threads, 4096-token context).
• 3. Make API requests in English and Chinese with temperature 0, seed 42, 512-token output budget.
• 4. Verify outputs match expected solution (x=7) and substitution check.
• 5. Record LM Studio 0.4.24 session (Think option disabled) with 3x speed label.
• 6. Compare Unsloth API results with LM Studio screen recording.
• 7. Document prompt, settings, raw output for reproducibility.

Evidence

Both recorded runs produced x = 7, with a correct substitution check.

Source

The Unsloth backend used build b10840, four CPU threads and a 4096-token context.

Source