AI Agent Builders Tools
PRODUCT OVERVIEW
TI-ONE is Tencent Cloud's large-model training and inference platform built for AI engineers. It covers the full path from data preparation and dataset labeling to task-based modeling, cloud development machines, drag-and-drop workflow orchestration, model repositories, evaluation and online model services. A model marketplace offers one-click deployment of mainstream open models such as Hunyuan, DeepSeek, Qwen, GLM and Kimi, while built-in templates support one-click fine-tuning and reinforcement learning. GPU management includes virtualization down to 0.01-card slices, quota controls, tidal scheduling that mixes training and inference, and adaptation for NVIDIA and domestic accelerators. The self-developed TACO inference engine and PD-disaggregated deployment target lower latency and higher throughput. Billing combines cloud server subscription fees with TI-ONE subscription fees, and compute can be purchased monthly or on a pay-as-you-go, per-second basis.
Sign in to the Tencent Cloud console and open the TI-ONE v2 console to access data, training, model management and model service modules.
Choose a mainstream open model in the model marketplace and publish it as an inference service with recommended resources.
Use built-in fine-tuning templates for mainstream open models, or launch reinforcement learning training with the native Ray and veRL support.
Allocate GPU resources through resource groups, quotas and GPU virtualization slicing, and use tidal scheduling to share capacity between training and inference.
Configure automatic or manual scaling, multi-version gray releases with rollback, and review full service request records for production operations.
Quickly deploying new open-source large models as online services
Fine-tuning mainstream open models with ready-made templates
Accelerating inference for long-text and high-concurrency workloads
Sharing GPU clusters across teams with quotas and isolation
Running training at night and inference during the day on the same pool
Training and serving models on domestic accelerator cards
Evaluating and comparing multiple models before launch
Supporting embodied intelligence training and world-model simulation
Public cloud billing combines cloud server subscription fees with TI-ONE subscription fees. Compute can be purchased from TI-ONE monthly or pay-as-you-go by second, or reused from purchased CVM instances.
Yes. When selecting compute from CVM, spot instances are supported and can offer lower compute prices.
Depending on the scenario, other cloud products such as COS, CFS, GooseFS, TCR, TCOP, CLS and CLB may be activated and billed separately.
The marketplace includes mainstream open models such as Tencent Hunyuan, DeepSeek, Qwen, GLM and Kimi, and supports one-click publishing as an inference service.
It provides built-in fine-tuning templates for mainstream open models and natively supports the Ray distributed framework with built-in veRL for one-click reinforcement learning training.
It separates Prefill and Decode roles so each can be configured with its own compute type and custom image, and is documented to improve first-token and inter-packet latency.
Yes. It unifies management of NVIDIA and domestic chips such as Hygon, Kunlunxin, Tianshu and Zixiao, with virtualization slicing supported for domestic cards.
It offers resource-group quotas, GPU virtualization slicing down to 0.01 card, and tidal scheduling with priority preemption and idle resource lending to share capacity between training and inference.
TI-ONE public cloud uses a combined billing model of cloud server subscription fees plus TI-ONE subscription fees. Compute can be bought from TI-ONE monthly or pay-as-you-go billed by second, or reused from purchased CVM instances including spot instances. Other cloud products such as storage, image registry, monitoring, logging and load balancing may be billed separately.
Compute resources purchased from the TI-ONE platform on a monthly subscription, described as convenient with a short subscription path.
Compute resources purchased from TI-ONE on demand, billed by second and stoppable at any time.
Reuse already purchased CVM machines, supporting monthly, pay-as-you-go and spot instance options.
COMMUNITY REVIEWS
No reviews yet.
Open-source model community and Model-as-a-Service platform for hosting, evaluating, fine-tuning, and deploying AI models, datasets, and applications.
