Together AI

Together AI is an AI Native Cloud founded in San Francisco in 2022 that turns open-weight models into production-grade, fine-tunable infrastructure. It does not train foundation models of its own: it hosts 200+ open weights from 38 model labs and puts its research into inference engines, kernels, and GPU scheduling, serving developers and teams who have to answer for unit economics.

Core capabilities

• Serverless inference: OpenAI-compatible, so you point base_url at it and run 200+ models with zero migration cost
• Batch inference: submit offline jobs asynchronously at about half price, ideal for evals and data generation
• Provisioned Throughput (PTU): $0.05 per PTU per minute with a 1-month minimum, buying reserved throughput and a 99% SLA
• Dedicated instances and GPU clusters: single-tenant H100/B200 up to bare metal, preemptible H100 from $1.99 per GPU-hour
• Fine-tuning and post-training: SFT, DPO, and custom reinforcement learning in one place, with SFT from $0.34 per 1M training tokens
• Systems research shipped in the product: FlashAttention-4, ThunderKittens, and ATLAS run in the platform, so the same model gets faster and cheaper

Where it fits Product teams running DeepSeek, Qwen, GLM, Kimi, or Llama in production; engineering leads whose monthly bills run from thousands to tens of thousands of dollars and who need the unit cost to add up; regulated industries that need a fine-tune-to-deploy loop plus data compliance; teams that need a big pool of GPUs for a short burst of training and RL rollouts.

What sets it apart OpenRouter is a resale layer with no GPUs of its own; Fireworks has lower latency but costs more, and Groq is fast but carries only a dozen or so models. Together AI's difference is owned compute plus its own kernels, with inference, fine-tuning, and clusters on one bill, publicly listed prices you can recalculate, and the backing of $1.15B in bookings plus a $240M long-term deal with IBM.

Editorial take For open-model hosting it is one of the two or three vendors most worth listing as a primary provider. The community line is "fast, cheap, and you can push tokens through it with confidence"; the deductions are all about governance: no free trial and a $15 top-up to start, models retired without notice, and recurring complaints about rate limits and refunds, while flagship first-token latency trails Groq and Fireworks.