deliany.com

What It Is

Deliany is an Apple Silicon training cluster provider based in Austin, Texas, focused on fine-tuning open-weight large language models on MLX. It is not a general-purpose AI SaaS product; it is closer to training infrastructure and managed fine-tuning for developers and AI teams. Its pitch is to use Apple Silicon’s large unified memory to handle memory-constrained training workloads for 32B- and 70B-class models.

Core Capabilities

Its main use cases include LoRA/QLoRA adapter training, full supervised fine-tuning (SFT), and preference alignment methods such as DPO, ORPO, and KTO. The site explicitly states that adapters can be trained on Llama 3.3 70B, Mistral Large, and Qwen 2.5 72B, while full-parameter SFT is available for Mistral 7B, Qwen 2.5 7B, Phi-4, and Gemma 2 9B. It supports ecosystems such as MLX, PyTorch, Hugging Face, Weights & Biases, and Ollama, but the emphasis is on MLX-native recipes.

Usage and Pricing

Deliany offers two modes: DIY node rental, where you SSH into a node and run your own MLX scripts; or a managed mode, where you submit a JSONL dataset and training configuration, and Deliany runs, monitors, evaluates, and delivers the adapter plus a summary report. Pricing is available hourly, monthly, or per job, with managed tasks quoted at a fixed per-job rate. It also highlights no spot market, no minimum commitment, and no surprise bills. However, the site does not publish specific prices, so actual cost-effectiveness still needs to be confirmed by requesting a quote.

Pros and Cons

The main strengths are its clear positioning and suitability for large-model fine-tuning tasks with high VRAM/memory pressure. US data residency, physical security, dual-carrier networking, and engineer support via Slack Connect are also clearly stated. The drawbacks are the lack of public pricing, free trial, Chinese-language service, and formal API information. The Apple Silicon/MLX ecosystem is also narrower than NVIDIA/CUDA, so some training workloads may not be a good fit.

Who It’s For and Access from China

Deliany is a fit for teams that already have datasets and want to fine-tune open-source LLMs, especially customers looking to try a non-NVIDIA path while requiring US data residency and engineering support. Access from mainland China, payment methods, and Chinese-language customer support are not disclosed, so they should be treated as unknown. If connectivity or compliance is a concern, consider comparing alternatives such as RunPod, Lambda Labs, and Modal, or domestic options like Alibaba Cloud PAI, Tencent Cloud TI, and Volcano Engine.