Alternatives to Together AI

Fireworks AI runs open-model inference with per-token pricing, model routing, and fine-tuning on one platform.

Serverless APIs for 200+ open-weight LLM, image, audio, video and vision models, plus GPU cloud and an agent sandbox.

Serverless AI inference on 100+ open models via an OpenAI-compatible API, billed per token with no GPU overhead.

Run fast, production-grade GPU inference on dedicated GPUs or token-priced Model APIs — with per-minute billing.

Rent per-second GPU pods, autoscaling serverless inference endpoints and multi-node clusters on Runpod's AI developer cloud.

Lilac turns idle GPU clusters into a white-labeled neocloud with spot H100 leases and serverless inference.

vLLM turns open-weight LLMs into high-throughput, OpenAI-compatible inference endpoints on your own GPUs.

FuriosaAI's RNGD inference accelerators and NXT RNGD Server serve LLMs in 3 kW air-cooled racks via OpenAI-compatible APIs.

Flat-rate subscription offering unlimited inference on open models GLM, Qwen, DeepSeek, and Kimi, tuned continuously for your traffic type.

Mesh LLM splits large open-weight models into layer ranges across your own GPUs and serves them through one OpenAI-compatible endpoint at localhost:9337.

SambaNova Cloud delivers fast open-weight LLM inference — MiniMax M3.1 Flash, DeepSeek-V3.1 and gpt-oss-120b on custom RDU silicon.

Open-source BentoML toolkit for packaging and serving diffusion models like Stable Diffusion and Flux as scalable APIs.

Groq delivers fast LLM inference on its own LPU and LPX silicon through an OpenAI-compatible API.

Runware is one inference API for 400K+ image, video, audio, LLM and 3D models on a single endpoint.

Managed Ray platform for distributed model training, multimodal data curation, and batch embedding generation on your own GPUs.

MAX Engine serves open models through one OpenAI-compatible endpoint across NVIDIA, AMD, Apple silicon, CPUs, and ASICs.

Salad Cloud rents idle consumer GPUs from $0.015 per GPU-hour, billed per second, for retryable inference and batch work.

TensorRT-LLM is NVIDIA's open-source inference library for high-throughput LLM and visual generation serving on NVIDIA GPUs.

Deploy your own ML models on serverless Nvidia GPUs, billed per second with scale-to-zero.

OpenAI-compatible multimodal inference API on dedicated NVIDIA GPUs, from $2.60/GPU-hour

ViewComfy turns ComfyUI workflows into no-code web apps and autoscaling serverless APIs, billed per active GPU second.

MatrixHub is a self-hosted AI model registry that gives vLLM and SGLang clusters a private Hugging Face replacement.

Predibase was a LoRA fine-tuning and multi-LoRA inference platform for open-source LLMs — now under Rubrik.

Open-source KV cache layer for LLM inference that skips redundant prefill work

European GPU cloud with self-service NVIDIA GB300, B300 and H200 instances plus serverless GPU inference

KubeAI is a Kubernetes operator for self-hosted LLM, embedding, reranking, and speech-to-text inference with an OpenAI-compatible API.

Crusoe Cloud GPU instances, serverless inference, and fine-tuning priced for AI teams that want orchestration included

Self-hosted control plane that turns your own GPUs into OpenAI- and Anthropic-compatible model endpoints

Run Llama 3.1 405B and other huge language models at home over a BitTorrent-style decentralized inference network.

Cocoon is Telegram's decentralized confidential AI inference network on TON, paying GPU owners in TON to serve models inside Intel TDX enclaves.