27B
General multimodal model
Qwen3.6 27B
Chat, document and image analysis, coding and long-context tasks.
QuantizationGGUF Q4_K_M or Q5_K_Mabout 17–22 GB of weights
ConfigurationOne ASUS GX10 node
GPU COMPUTE FOR LLM & AI$5.07from$1.07/ hour
ASUS Ascent GX10 rental
ASUS GX10 has 128 GB of memory shared by the CPU and GPU. Account for model weights, context, the operating system and working data within this capacity.
ASUS GX10 with 128 GB of unified memory
128 GB of unified memory for model weights, context and LLM experiments.
128GBshared CPU and GPU memory
NVIDIA GB10Grace Blackwell · ARM64
Offer ends in—d—h—m—s
MODEL CATALOG
Choose a model for your workload
The catalog lists workloads, quantizations and weight sizes. Six models are preinstalled on ASUS GX10.
Configuration: one ASUS GX10
19 models · 6 preinstalled
35B · 3B активных
Fast MoE model for coding and analysis
Qwen3.6 35B-A3B
For coding, chat and image analysis. The MoE architecture activates only part of its parameters per token.
QuantizationGGUF Q4_K_M or Q5_K_Mabout 22–28 GB of weights
ConfigurationOne ASUS GX10 node
80B
Code, repositories and agents
Qwen3-Coder-Next 80B
A specialist model for codebases, debugging, code generation and tool-using workflows.
QuantizationGGUF Q3_K_M or Q4_K_Mabout 36–48 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
117B · 5.1B активных
Large open model for demanding prompts
gpt-oss-120b
OpenAI’s MoE model with native MXFP4 weights for reasoning, coding and local assistant workloads.
QuantizationOfficial MXFP4 weightsabout 63 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
30B · 3B активных
Image and video understanding
Qwen3-VL 30B-A3B
Answers questions about screenshots, photos, video and visually rich documents.
QuantizationGGUF Q4_K_Mabout 18–21 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
32BPreinstalled
Chat and writing
Qwen3 32B
Preinstalled on ASUS GX10. Works in English and Russian across writing, translation and document tasks.
QuantizationOllama Q4_K_Mabout 20 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
30B · 3B активныхPreinstalled
Development and coding agents
Qwen3-Coder 30B-A3B
Preinstalled on ASUS GX10. Writes and explains code, helps debug apps and work with tools.
QuantizationOllama Q4_K_Mabout 19 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
32BPreinstalled
Reasoning, maths and logic
DeepSeek-R1 32B
Preinstalled DeepSeek-R1 distillation based on Qwen 32B for multi-step problems and analysis.
QuantizationOllama Q4_K_Mabout 20 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
27BPreinstalled
Images and documents
Gemma 3 27B
Preinstalled on ASUS GX10. Understands text and images, including screenshots, diagrams and document pages.
QuantizationOllama Q4_K_Mabout 17 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
70BPreinstalled
Chat and long documents
Llama 3.3 70B
Preinstalled on ASUS GX10 for detailed answers, summarization and conversational applications.
QuantizationOllama Q4_K_Mabout 43 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
8BPreinstalled
Uncensored chat and creative writing
Dolphin 3.0 8B
Preinstalled uncensored model based on Llama 3.1 for open-ended chat and creative writing.
QuantizationOllama Q4_K_Mabout 4.9 GB of weights
ConfigurationOne ASUS GX10 node
Uncensored: yesUncensored is an explicit characteristic of this Dolphin release.
5,6B
Image generation and editing
FLUX.3 Image Klein
Black Forest Labs open weights for prompt-based image generation and editing an input image.
QuantizationFP8 or NVFP4about 6–12 GB of weights; pipeline size varies by build
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
20B
Image generation and in-image text
Qwen-Image-2512
Qwen image generation and editing model, including compositions with embedded text.
QuantizationFP8 or 4-bit buildsize depends on format; include the VAE and text encoder
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
5B
Video from text or image
Wan2.2 TI2V 5B
One model supports text-to-video generation and animation of an input image.
QuantizationFP8 or 8-bit buildfull pipeline is about 34 GB of files
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
14B · MoE
Text-to-video generation
Wan2.2 T2V A14B
A larger Wan MoE model for text-to-video. We check the full pipeline and speed for the requested resolution.
QuantizationFP8 or GGUF Q4leave room for the text encoder, VAE and runtime buffers
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
7B база + LoRA
Financial news and text
FinGPT · Sentiment
For financial sentiment classification and analysis of news and reports. It does not provide a ready-made trading strategy.
QuantizationQ4_K_M + financial LoRAabout 5–7 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo uncensored build is selected for this financial model.
200M
Time-series forecasting
TimesFM 2.5
A specialist time-series forecasting model for experiments with historical market data and backtests.
QuantizationBF16 or FP16less than 1 GB of weights
ConfigurationOne ASUS GX10 node
Not applicableThis is not a chat model; an uncensored version does not apply.
1,7B
Speech synthesis and voice design
Qwen3-TTS 1.7B
Generates speech from text and supports voice design. We check the selected voice and mode for Russian in advance.
QuantizationBF16 / FP16about 3–4 GB of weights plus audio components
ConfigurationOne ASUS GX10 node
Not applicableThis is not a chat model; an uncensored version does not apply.
141B · MoE
MoE model for chat and code
Mixtral 8×22B
A large Mistral model for text and coding. It can run quantized on one node or be split across two.
QuantizationGGUF Q3_K_M / Q4_K_Mroughly 65–90 GB of weights
ConfigurationOne ASUS GX10 node
No verified version selectedNo verified uncensored build is selected for this catalog.
NVIDIA GB10 · ASUS ASCENT
Rent one ASUS GX10 with 128 GB of unified memory
One ASUS Ascent GX10 has 128 GB of LPDDR5x memory shared by the CPU, GPU and operating system. Include model weights, KV cache, context and working data when estimating memory use. The system can run quantized LLM inference, RAG, AI agents and fine-tuning tests when the full workload fits in available memory.
ASUS GX10 uses the NVIDIA GB10 Grace Blackwell platform and an ARM64 CPU. Containers, libraries and CUDA extensions must support this architecture. Share your model, quantization, context length and software versions so we can check compatibility and confirm the rate, availability and rental period.
Compare with NVIDIA DGX SparkRENTAL COST
ASUS GX10 rental rate
One hourly rate applies to the selected rental duration. The calculator multiplies it by the full rental period.
The rate is for one ASUS Ascent GX10 with 128 GB of memory shared by the CPU and GPU.
ASUS Ascent GX10$5.07from$1.07 / hour
Offer ends in—d—h—m—s
Rental 1–5 h$5.07$3.02/ h
Rental 6–11 h$5.07$2.29/ h
Rental 12–23 h$5.07$1.99/ h
Rental 24–168 h$5.07$1.69/ h
Rental > 7 days$5.07$1.07/ h
Estimated budget$18.35Rate for selected duration: $2.29 / h
PRICE AND MEMORY COMPARISONASUS GX10 · 128 GB
ASUS GX10 · 128 GB
Rates and memory
The table lists the hourly rate and memory capacity for each configuration.
GX10 uses memory shared by the CPU and GPU. RTX and A100 use dedicated GPU memory.
from $1.07/ч
The “from” rate applies to rentals of 169 hours or more.
Offer ends in—d—h—m—s
128 GB of unified memory per node
The CPU, GPU, operating system and applications share LPDDR5x. The operating system and runtime use part of the memory.
How to use two nodes
Run independent workloads at the same time or distribute one model when your software supports multi-node execution.
| Configuration | Memory | Per hour |
|---|---|---|
| ASUS Ascent GX10NVIDIA GB10 · Blackwell | 128 GBCPU + GPU | Offer ends in—d—h—m—s |
| 2 × ASUS Ascent GX10NVIDIA GB10 · Blackwell | 2 × 128 GBunified per node | Offer ends in—d—h—m—s |
| RTX 3090 | 24 GBVRAM | $0.40/ hour |
| RTX 4090 | 24 GBVRAM | $0.712/ hour |
| RTX 5090 | 32 GBVRAM | $0.952/ hour |
| A100 SXM 80 GB | 80 GBVRAM | $1.432/ hour |
| H100 SXM 80 GB | 80 GBVRAM | $3.192/ hour |
Workloads for this GPU configuration
Run an LLM on one node
Run the model and working data on one ASUS GX10 if they fit in available memory. Compare model sizes and quantization formats without setting up a cluster.
Context for RAG and agents
Use the memory budget for KV cache, documents in context and application data. Match conversation length and concurrent requests to your service.
Fine-tuning with your data
Test LoRA or QLoRA on your dataset. Memory for weights, activations and buffers depends on the method, model and batch size.
MEMORY ESTIMATE
Estimate memory for an LLM
Estimate model weight memory first. Then add the KV cache, runtime, operating system and working buffers.
| Parameters | 4 bit | 8 bit | 16 bit |
|---|---|---|---|
| 8B | 4 | 8 | 16 |
| 14B | 7 | 14 | 28 |
| 32B | 16 | 32 | 64 |
| 70B | 35 | 70 | 140 |
| 120B | 60 | 120 | 240 |
Calculation: parameter count × bit width / 8. Excludes quantization metadata, context and runtime overhead.
70B at 4 bits: about 35 GB of weights
A 24 GB card cannot hold these weights entirely in VRAM, even before adding context. GX10 offers a larger shared 128 GB CPU and GPU memory budget. Context length and concurrent requests also determine the final footprint.
When to choose a cluster
Choose the two-system ASUS GX10 cluster for models that need more than 128 GB or for parallel workloads.
All models for ASUS GX10TECHNICAL DETAILS
ASUS GX10 specifications
Each node has 128 GB of memory shared by the CPU and GPU. The system is built on NVIDIA GB10 Grace Blackwell.
ASUS technical specifications- Devices
- ASUS Ascent GX10
- Compute platform
- NVIDIA GB10 Grace Blackwell
- GPU
- NVIDIA Blackwell, integrated in GB10
- CPU
- 20 ARM64 cores · Arm v9.2-A
- Memory
- 128 GB, shared by CPU and GPU
- Memory type
- LPDDR5x · unified system memory
- Network adapter
- NVIDIA ConnectX-7
- OS platform
- NVIDIA DGX OS · based on Ubuntu Linux
RENTAL STEPS
Prepare for the first runHow to rent ASUS GX10
- 01
Describe your workload
Share your model, weight format, engine, context length and preferred dates. For fine-tuning, include the method and dataset size.
- 02
Agree on the setup and terms
Discuss ARM64 compatibility, access, storage, the final price and availability. A request becomes a rental once the terms are agreed.
- 03
Start with a representative run
Connect using the agreed method and run a small example. Measure memory and speed at your target context length before starting the main workload.
Software compatibility
- CPU, GPU and the system share 128 GB LPDDR5x. It is not 128 GB of dedicated VRAM.
- Check ARM64 container images and compatibility of custom CUDA extensions.
- Weights, KV cache, runtime and working buffers must fit together.
Rent ASUS Ascent GX10
Provide your model, software environment and rental period. We will confirm the rate, availability and connection method before the rental starts.
Offer ends in—d—h—m—s
FAQ
ASUS GX10 rental questions
How much does ASUS Ascent GX10 rental cost?
The current rate appears on this page. Budget for runtime as well as the hourly price: more memory can help a workload fit, but does not determine its speed. The final price, rental duration and resources are agreed on request.
Is 128 GB of unified memory the same as VRAM?
It is LPDDR5x system memory shared by the CPU and GPU, rather than dedicated graphics memory. The OS, runtime and buffers use part of it. Model weights and KV cache must fit together. Its performance is not directly interchangeable with GDDR or HBM on a discrete GPU.
Can I run a 70B-parameter LLM?
A quantized 70B model may fit, depending on the format, context and engine. FP16/BF16 weights alone need roughly 140 GB, so do not expect a fully memory-resident unquantized 70B model in 128 GB. At 4 bits, the basic weight estimate is about 35 GB, plus metadata, KV cache and runtime. Share the exact model for an assessment.
Can I use Ollama, PyTorch and Docker?
These tools have GX10 deployment paths, but compatibility depends on the version and build. A Docker image must support linux/arm64; PyTorch, CUDA libraries and the inference engine must support NVIDIA GB10/Blackwell. An x86-64 image may need rebuilding. Share exact versions and any custom CUDA extensions before agreeing on the environment.
What is the real generation speed in tokens per second?
We have not yet published our own TPS measurements for the offered GX10. Speed depends on the model, quantization, input length, context, engine and concurrency. Compare time to first token and subsequent generation speed separately using the same workload.
When is one GX10 enough, and when do I need a cluster?
Start with one node if the model, context and working data fit and performance meets your needs. A second node supports parallel jobs or a compatible distributed setup. It does not automatically remove a single process’s memory limits. Both configurations use the same pool of two devices, so their availability is linked.
How is ASUS Ascent GX10 different from NVIDIA DGX Spark?
The offered node is ASUS Ascent GX10, rather than the NVIDIA DGX Spark device. Both are based on NVIDIA GB10 Grace Blackwell and related software; enclosure, cooling and configuration options depend on the manufacturer. DGX OS is NVIDIA’s operating-system name and does not change who makes GX10.
What should I prepare for a request and first run?
Include the model link, quantization format, engine and container versions, context length and concurrency. For fine-tuning, add the method, batch size and dataset size. Share dates, duration, and access and retention requirements. Prepare a small first-run example; keep secrets and private keys out of the request.
GPU selection and setup guides
HARDWARE
ASUS GX10: 128 GB for LLMs
How to account for unified memory, context length and ARM64 compatibility.
CHOOSING A SETUPOne ASUS or two?
When one node is enough, and when parallel or distributed computing helps.
DOCUMENTATIONMatch a GPU to your workload
Memory, architecture and software compatibility: three things to start with.
