Why we track hosting plans and models separately
The same model can cost and behave very differently depending on the plan you run it on.
LLM Hosting Pro
News, reviews and resources on where to run large language models.
Latest news
All news →Per token or per GPU-hour? Choose the billing model that fits your traffic
Spiky, low-volume traffic favours per-token plans. Steady, high-volume traffic often doesn't.
Review: RunPod Serverless for open models
Flexible and inexpensive at low volume, but you own the container and the cold starts.
Review: Together AI Serverless Inference
A broad catalogue behind an OpenAI-compatible API makes this an easy place to start.
gpt-oss-120b is available on several serverless hosts
OpenAI's open-weight model can be called per token without running your own GPUs.
LLM hosting plans
Compare all plans →Serverless APITogether AI
Together AI Serverless Inference
Shared, pay-as-you-go endpoints for a large catalogue of open models behind one API.
PricingPay per token
Dedicated endpointTogether AI
Together AI Dedicated Endpoints
Reserved GPUs running the model you choose, for steady latency and throughput.
PricingPer GPU-hour
Serverless APIFireworks AI
Fireworks AI Serverless
Fast shared inference for popular open models, with structured output and function calling.
PricingPay per token
Dedicated endpointFireworks AI
Fireworks AI On-Demand Deployments
Private deployments on reserved GPUs with no per-token rate limits.
PricingPer GPU-hour, Per second of GPU time
Serverless APIGroq
GroqCloud Developer Tier
Very fast token generation on Groq's LPU hardware, with a free tier to start.
PricingPay per token
Serverless GPURunPod
RunPod Serverless
Run your own container or a ready-made vLLM worker on GPUs that scale with requests.
PricingPer second of GPU time
LLM models
All models →Open weightsMeta
Llama 3.3 70B Instruct
A general-purpose instruction-tuned model and a common default for open-weight chat workloads.
- Parameters
- 70B
- Context
- 128K tokens
Open weightsMeta
Llama 3.1 8B Instruct
A small, fast model for classification, extraction and high-volume, low-cost tasks.
- Parameters
- 8B
- Context
- 128K tokens
Open weightsAlibaba Cloud (Qwen team)
Qwen2.5 72B Instruct
A strong multilingual model with good coding and structured-output performance.
- Parameters
- 72B
- Context
- Up to 128K tokens
Open weightsDeepSeek
DeepSeek-V3
A large mixture-of-experts model that competes with closed frontier models on many benchmarks.
- Parameters
- 671B total, 37B active (mixture of experts)
- Context
- 128K tokens
Open weightsOpenAI
gpt-oss-120b
OpenAI's open-weight reasoning model, released under a permissive licence.
- Parameters
- 117B total, about 5B active (mixture of experts)
- Context
- 128K tokens
Resources
Guides and tools →Choosing how to host an LLM
Managed API, serverless GPU or your own GPU box? A plain-English way to pick the right hosting model for your project.
Estimating LLM hosting costs
How to compare per-token pricing with GPU hours, and the hidden costs that catch teams out.
Self-hosting an open model with vLLM
A first-deployment checklist for serving an open-weight model on your own GPU.