Models and platforms
11 totalOpen models, model hosting, inference APIs and leaderboards.
This category is not an end-user interface. It is for people building products or deploying privately: downloadable open weights, local runtimes you can start with one command, token-priced inference APIs, and leaderboards for comparing models.
Decide one thing first: must the data stay on your machine? If so, choose a runtime plus open weights. If you only need to connect a model to your app, an inference API is easier to maintain.
Choose by your constraints:
- Offline or private deployment: a local runtime plus open weights of the right size. VRAM determines the parameter count you can run.
- Connect to an app without managing machines: use an inference API and compare latency and price per million tokens for comparable models. Check the official pricing page for numbers.
- Weights and datasets: go to an open hosting community, and read the license of each weight separately.
- Side-by-side evaluation: use public arena leaderboards, and expect rankings to shift as models are updated.
- Mainland China filing and compliance: prefer mainland inference platforms and pair them with your own content and audit logging.
Tencent Hunyuan
Tencent's entry point for trying its own models and building on them
FreeChina mainland
Website ↗
OllamaHOT
A command-line tool for running open models locally
FreeInternationalOpen source
Website ↗
LM Studio
Local model running and management with a graphical interface
FreeInternationalOpen source
Website ↗
llama.cpp
An open-source C++ inference library and quantization toolkit
FreeInternationalOpen source
Website ↗
vLLM
An open-source high-throughput inference framework for serving
FreeInternationalOpen source
Website ↗
SiliconFlow
A domestic model inference cloud with pay-as-you-go APIs
Free tierChina mainlandAPI available
Website ↗
OpenRouter
One API to switch between many model providers
Free tierInternationalAPI available
Website ↗
Hugging FaceHFHOT
A hosting community for models and datasets
FreeInternationalAPI available
Website ↗
No tools match in this category.Back to all →
Frequently asked questions
What GPU do I need to run a large model locally?
It depends on the quantized weight size. 7B to 8B quantized weights run on consumer cards with 8GB of VRAM; 14B to 32B need more VRAM or multiple cards. Context length also consumes VRAM, so do not estimate from parameter count alone.
What are the licensing traps with open models?
Different versions of the same model may use different licenses. Some require a separate license above a monthly-user or revenue threshold, and some forbid specific uses. Read the license on the weights page before integrating into a commercial product.
Can leaderboards guide model selection?
They show relative strength, not business fitness. Prompt distribution, judge models and battle samples all affect ranking. Testing ten real tasks of your own is more reliable than any leaderboard.
Are APIs from different model vendors interchangeable?
Most offer a form compatible with common APIs. Switching costs lie in parameter differences, rate limits and billing conventions. Before launch, send the same request and verify response structure and error codes.
Browse another use case
Back to the AI directory; for tools that run locally in your browser, see U2 online tools.