Uncensored LLM hosting on ASUS GX10

What uncensored and abliterated mean

Uncensored commonly describes models with modified training or refusal behavior. The term is an author’s description, so compare specific versions on your own prompts. Dolphin 3.0, for example, is described by its developers as a model whose behavior the operator steers through the system prompt.

Abliterated refers to a modification of model weights intended to reduce refusals. The author of huihui-ai/Qwen3-32B-abliterated explicitly identifies this method in its model card. Evaluate the changed behavior alongside answer accuracy and the output format your application needs.

Choose weights and check the license

Specify the full repository ID and revision. The dphn/Dolphin3.0-Llama3.1-8B card lists the Llama 3.1 license; huihui-ai/Qwen3-32B-abliterated lists Apache 2.0 in its metadata. Read the license file and conditions attached to the chosen repository, including the base model’s terms, before use.

Other candidates to evaluate include dphn/Dolphin-Mistral-24B-Venice-Edition, whose card lists Apache 2.0, and huihui-ai/Llama-3.3-70B-Instruct-abliterated, which lists Llama 3.3. A rough estimate for 70B weights at 16 bits is 140 GB before overhead. A single 128 GB GX10 therefore needs a suitable quantization option and a verified full memory budget.

Next, select a weight format and quantization that suit the inference engine. Record the file name or digest, chat template and context settings. The uncensored label does not specify the model format, memory requirements or redistribution terms.

Validate the stack on ARM64 and GB10

GX10 uses an ARM64 CPU. For an existing web stack, check the container architecture, dependencies and GB10 support in the chosen engine. Ollama provides a Linux ARM64 package and llama.cpp documents an HTTP server; the exact build, model and web interface still need validation on the device.

When testing the application, check system-prompt handling, streamed responses and API error handling. Configure access, authentication and conversation storage for your deployment. A hardware listing does not mean this service is already deployed at a ready-to-use endpoint.

Budget memory and rental time

The offered GX10 systems each have 128 GB of unified memory. Weights share it with the system, runtime and KV cache; longer contexts and concurrent conversations change memory use. Choose one node for the initial run and verify consumption with a representative workload.

A GX10 pair consists of two independent 128 GB nodes. It can host separate environments; distributing one model requires a compatible runtime and an interconnect configuration. Include weight downloads, setup and testing in the rental period. Rates and the calculator are on the configuration page; availability and dates require separate confirmation.