DEV Community

SparkLLM
SparkLLM

Posted on

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

#ai #machinelearning #tutorial #opensource

Watch the short

Spark-X2.5 can run locally as GGUF with the Unsloth llama.cpp backend and a compatible LM Studio runtime. This demo brings those two routes together: independent checks with an Unsloth-built backend, followed by real screen recordings in LM Studio.

One equation, two languages

We used the official Spark-X2.5-1.7B Q4_K_M GGUF and asked the same mathematical question in Chinese and English:

Solve 3(x - 2) + 4 = 19.
Briefly show the steps and verify the result by substitution.
Enter fullscreen mode Exit fullscreen mode

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

3x - 6 + 4 = 19
3x - 2 = 19
3x = 21
x = 7

3(7 - 2) + 4 = 3 * 5 + 4 = 19
Enter fullscreen mode Exit fullscreen mode

The footage shows the actual prompt entry, submission, generation and scrolling through the result. The screen-recording section is clearly labelled 3x speed. It is continuous recorded footage, not a sequence of screenshots standing in for a run.

What was tested

The recorded client is LM Studio 0.4.24, with the model's Think option disabled. We also made separate Chinese and English API requests to the Unsloth-built llama.cpp Windows CPU backend, using the same official model file. Those checks passed this example as well.

The Unsloth backend used build b10840, four CPU threads and a 4096-token context. The preflight requests used thinking disabled, temperature 0, seed 42 and a 512-token output budget. The official GGUF file was checked against its Hugging Face LFS SHA-256 metadata.

These are separate runs. The LM Studio footage is not an Unsloth Studio interface recording, and API timings should not be presented as timings for the recorded UI session. This video is an example demonstration, not a backend speed comparison.

The selected equation passed preflight testing before filming. Earlier attempts on a different fraction equation, including wrong or truncated outputs, were retained. Passing this example does not establish general mathematical accuracy.

Try your own local task

  1. Open the Spark-X2.5 project and its Model Downloads section.
  2. Choose an official GGUF checkpoint, such as Spark-X2.5-1.7B-GGUF.
  3. Use a runtime containing Spark2_5 architecture support. For Unsloth, see the verified backend release; for Studio model loading, follow the current guide.
  4. Start with a task whose answer you can independently check. Preserve the model revision, prompt, settings and raw output.

From math exercises to writing, coding tools and agent experiments, we welcome useful tests and creative applications. Download, explore, and share your feedback, experiences and creations.

Spark-X2.5 · Official GGUF · Unsloth · Watch the demo

Music: Open Horizons, an original synthesized instrumental made for this episode.

Top comments (0)