Together AI Dedicated Inference
together.ai
- Web
- API
6.5 Club score Visit together.ai
Plans
Every plan and priceDedicated Inference — reserved capacity
—
Dedicated Inference — NVIDIA HGX H100 (on-demand)
$5.49 per contact
Dedicated Inference — NVIDIA HGX B200 (on-demand)
$8.99 per contact
What the maker publishes
- API compatibility
- Dedicated and serverless inference use the same OpenAI-compatible API, so a serverless prototype can move to a dedicated endpoint by changing the endpoint. source
- Autoscaling and routing
- The product page lists SLO-driven autoscaling based on time to first token, latency, and throughput, along with least-loaded, session-affinity, and prefix-cache-aware routing. source
- Best suited for
- Together lists predictable or steady traffic, latency-sensitive applications, and high-throughput production workloads as best suited to Dedicated Model Inference. source
- Company mission
- Together AI describes its mission as designing a full-stack AI platform powered by systems research to help teams ship faster, scale reliably, and achieve superior unit economics. source
- Compliance and data residency
- Together says customers can choose a deployment region for data residency and identifies SOC 2 Type II and ISO 27001 as supporting certifications. source
- Custom models
- Customers can deploy custom or fine-tuned models from Hugging Face or their own S3 storage through the UI or CLI. source
- Deployment
- Customers select a target model and hardware configuration to deploy a production-ready endpoint, and Together says it can be live in minutes without deep infrastructure expertise. source
- Founded
- 2022 source
- Headquarters
- San Francisco, California, United States source
- Pricing basis
- Dedicated Inference is billed per GPU-hour based on GPU type and replica count; the pricing page lists on-demand rates for HGX H100 and HGX B200, while reserved rates require contacting sales. source
- Rollouts
- Customers can test updated deployments with shadow traffic or A/B testing, and Together says it manages the rollout without taking the serving endpoint down. source
- Security and privacy
- Dedicated deployments use single-tenant, isolated GPUs; Together says traffic and data are not shared and that it does not train on customer prompts or model weights. source
- Service
- Dedicated inference runs a model on GPUs reserved exclusively for a customer's workload, providing consistent latency, predictable cost, and control over the model and hardware. source
- Support
- The product FAQ directs users to contact Together's team for help, and the product page offers a Talk to Sales option. source
Checked against the maker's pages on 8 Oct 2026.