Skip to content

Together AI

Together AI is a full-stack cloud platform for running, fine-tuning, and deploying open-source AI models, aimed at engineering teams that need production inference without managing their own hardware.

Free TierModel APIs
190

Overview

Together AI is a cloud platform for running, fine-tuning, and deploying open-source AI models, built for engineering teams that want production-grade inference without operating their own GPU fleet. It combines a serverless model catalog, batch processing, dedicated endpoints, and GPU clusters under one roof, and it positions itself as infrastructure for companies building AI products rather than as an end-user application.

A Serverless Catalog Spanning Modalities

The platform's entry point is on-demand inference across text, image, video, code, and voice models. Teams call hosted open-source models through an API and skip provisioning entirely. The catalog is broad, ranging from small, inexpensive models to large frontier-class open-weight systems, and the pricing page lists per-token rates for each. For teams evaluating several models before committing, this removes the usual setup work of standing up a serving stack.

Batch Processing and Committed Capacity

Beyond interactive requests, Together AI offers asynchronous batch processing for large workloads, with the ability to scale to 30 billion tokens per model on either serverless models or private deployments. For steadier production traffic, committed inference capacity adds token-based pricing, reserved throughput, and a 99% uptime SLA. The company describes the API as drop-in compatible, which lowers the cost of migrating an existing integration.

Dedicated Endpoints and GPU Clusters

Teams that need more control can deploy models on dedicated infrastructure. Together AI also sells GPU clusters, including NVIDIA GB200 NVL72 and HGX B200 systems, with a configuration aimed at generative media workloads such as video, audio, and image models. The company holds NVIDIA preferred partner status, and its published customer work with Cursor describes porting the inference stack to ARM hosts, writing custom kernels for Blackwell Tensor Cores, and designing parallelism meshes across the 72-GPU NVL72 topology.

From New Weights to a Test Endpoint

The Cursor case study is the most concrete public description of the platform's workflow. Cursor's research team trains coding models internally and produces candidate weights; Together AI quantizes them using NVIDIA TensorRT-LLM and NVFP4, validates the result, and stands up a production-like test endpoint within days. Cursor then runs internal evaluations and staged A/B tests before cutover. That pipeline matters for any team that iterates on model weights frequently and needs a fast, repeatable path to serving.

Fine-Tuning and Post-Training

Alongside inference, the platform covers fine-tuning and post-training, so teams can adapt foundation models on managed infrastructure instead of assembling their own training environment. This fits organizations that want to keep model customization and serving with a single provider rather than splitting them across vendors.

Enterprise Support and Compliance

Together AI offers forward-deployed engineers for enterprise customers, drawn from the team that supports large inference deployments. The company reports serving more than 400 trillion tokens per month across customers that include Cursor, Zoom, Salesforce, and Mozilla. Compliance credentials include SOC 2 Type II and ISO 27001:2022. A published enterprise testimonial cites a 2x reduction in time to first token and roughly one-third lower costs, though that figure comes from a single customer and should not be generalized.

Where It Fits

The platform suits teams that have settled on open-source or open-weight models and need reliable serving at scale. Smaller projects can start on serverless inference and move to dedicated endpoints as traffic grows. The main trade-off is complexity: the catalog is large, pricing varies by model, and dedicated capacity and enterprise support involve a sales conversation. Teams that want a single managed vendor for inference, fine-tuning, and GPU capacity will find the breadth useful; those with narrow needs may prefer a simpler, single-model provider.

Core features

  • Serverless inference for open-source text, image, video, code, and voice models
  • Asynchronous batch processing for large workloads, scaling to 30 billion tokens per model
  • Committed inference capacity with token-based pricing, reserved throughput, and a 99% uptime SLA
  • Drop-in API compatibility for production workloads
  • Dedicated endpoints on dedicated infrastructure for teams needing speed and control
  • Fine-tuning and post-training services
  • GPU clusters, including NVIDIA GB200 NVL72 and HGX B200, for generative media and custom deployments
  • SOC 2 Type II and ISO 27001:2022 compliance, plus NVIDIA preferred partner status

Use cases

  1. 1Serving open-source LLMs through an API without provisioning GPUs
  2. 2Running large asynchronous batch jobs such as synthetic data generation
  3. 3Fine-tuning or post-training foundation models on managed infrastructure
  4. 4Deploying video, audio, and image generation models on GPU clusters
  5. 5Moving trained model weights to a production-like test endpoint for evaluation
  6. 6Reserving dedicated capacity for latency-sensitive production applications

Strengths and things to check

Strengths

  • Covers the full model lifecycle, from experimentation through dedicated production deployment
  • Serverless tier lets teams start without infrastructure commitments
  • Committed capacity includes a 99% uptime SLA for production workloads
  • Public customer work with Cursor describes a repeatable weights-to-endpoint pipeline and Blackwell-based serving
  • Compliance posture includes SOC 2 Type II and ISO 27001:2022

Things to check

  • Pricing is per-model and token-based, so costs vary widely by model choice
  • Dedicated and enterprise arrangements require working with a sales team
  • The catalog spans many models, which can complicate selection for smaller teams
  • Reserved capacity and SLAs apply to committed plans rather than the entry-level serverless tier

Pricing

Serverless inference is billed per 1 million tokens, with rates set per model; examples include Kimi K3 at $3.00 input and $15.00 output, GLM-5.3-Flash at $0.15 input and $0.50 output, and DeepSeek V4.1 Flash at $0.30 input and $1.20 output, with discounted cached-input rates on many models. Committed inference capacity uses token-based pricing with reserved throughput and a 99% uptime SLA. Prices are subject to change.

Basic information

Added
Oct 8, 2026
Platforms
Web
Chinese & English
Not marked
API
Not marked
Self-hosting
Not marked
Best for
AI engineering teams, model researchers, and product organizations that want to run open-source models in production without building and operating their own GPU infrastructure.
Information updated
Oct 16, 2026
Website status
Checking
Corrections · Privacy · Data requests?
Contact us

Features, pricing and availability may change. Always verify details on the official website. Maintained by steven · Last updated Oct 16, 2026 →

Related tools

All categories