Try an uncensored LLM before a longer server rental

Choose two models to compare

Consider Dolphin 3.0 Llama 3.1 8B for an initial run. To evaluate an abliterated version specifically, plan a comparison of Qwen/Qwen3-32B with huihui-ai/Qwen3-32B-abliterated. Dolphin’s card lists Llama 3.1, while both Qwen cards list Apache 2.0. Read the repository conditions and record revisions before downloading.

Uncensored describes modified model behavior, while abliterated identifies a weight modification intended to reduce refusals. Evaluate how well each candidate performs your task, such as answering from a document or returning valid JSON. The weights and runtime for this comparison still need validation on GX10.

Prepare a reproducible environment

ASUS GX10 needs ARM64-compatible dependencies and an engine build that supports GB10. Record the driver, runtime and web-application versions. Check the selected quantization, chat template and context limit. An Ollama ARM64 package or a llama.cpp HTTP server alone does not validate the complete application stack.

You can plan sequential runs of two models on one node. For concurrent independent tests, consider a GX10 pair if both systems are available. Each node has its own 128 GB of unified memory, with a separate budget for weights, KV cache and supporting processes.

Run prompts from your application

Use the same prompt set for both models: a short conversation, an answer grounded in your document, a code example or structured output. Apply each model’s correct chat template and record the system prompt and generation settings. Save responses alongside their inputs so you can inspect the differences.

Check factual errors, format compliance and the language your users need. In the web application, test streaming, cancellation and error handling. Measure time to first token and memory use in your own test; the result applies to that build and configuration rather than every model on GX10.

Allow time for the complete trial

Include downloading both models, setup, warm-up and saving results in the short rental period. Enter the planned duration in the ASUS GX10 calculator, since the hourly rate depends on rental length. Trial dates and availability of one node or a pair require separate confirmation.

After testing, retain the weight identifiers, runtime configuration and evaluation results. Use them to choose a model and budget for a longer rental, or adjust the next test. The site’s forms are currently demonstrations: submitting one neither starts a model nor confirms a reservation.