Together AI
Together AI is a San Francisco-based frontier-model cloud platform founded in 2022 with the slogan "decentralized AI," focused on an integrated open-source LLM training + inference + fine-tuning service. The company spun out of the early Apache TVM / OctoML team; its technical foundation is high-performance ML compilation and distributed training. In the generative AI wave, Together and Fireworks AI together form the "OpenAI alternative" core.
Key capabilities
Inference
- 100+ open-source models hosted (Llama / Qwen / DeepSeek / Mistral / Yi / Gemma etc.)
- Together Inference Engine, focused on low latency + low cost
- OpenAI-API compatible, smooth migration
- Function calling / JSON mode supported end-to-end
Training
- Together Training: distributed GPU clusters; LoRA / QLoRA / full parameter
- Released RedPajama open dataset (1.2T tokens), fueling open-source LLM training
- Collaborated with academic teams on open LLMs (e.g. StripedHyena / Yi)
Research
- Long-running investment in FlashAttention reproductions and extensions
- Together Kernel Collection: high-performance attention kernels
- Open-source LLMs like StripedHyena to compete with Llama / Mistral
Influence
- RedPajama dataset became a de facto standard for open-source LLM training
- Together with Fireworks AI pushed the "open models are commercially usable" frontier
- Serves flagship AI products like Character.ai, Perplexity, Salesforce Einstein
- Closed a $305M Series B in 2024 at a $3.3B valuation
Limitations
- Heavy feature overlap with Fireworks AI; price war compresses margins
- Training-cluster scale below hyperscalers; enterprise SLA trails Azure / Bedrock
- In-house models (StripedHyena) have limited mindshare, never reached the first tier of open-source models