llms.txt
Compare with other sites' llms.txt files →# Together AI
> Together AI is the AI Native Cloud: a full-stack platform for running, fine-tuning, and training
> open-source and frontier AI models in production, used by developers and AI-native companies for
> serverless and dedicated inference, model shaping, code sandboxes, and NVIDIA GPU clusters via an
> OpenAI-compatible API. This file covers www.together.ai — products, pricing, models, customers and
> company pages. API reference and how-to guides live on docs.together.ai, which publishes its own
> llms.txt and is the authoritative source for anything developer-facing.
The inference API base URL is https://api.together.ai/v1 (the older api.together.xyz host still
resolves). The model catalog changes most weeks, so read the live list rather than memorized model
names: GET https://api.together.ai/v1/models, or the serverless catalog page below. Model pages on this site are always /models/<slug>.
## Getting started
- [Quickstart](https://docs.together.ai/docs/quickstart): Make your first Together AI API request.
- [Docs index for LLMs](https://docs.together.ai/llms.txt): Auto-generated index of the full docs site.
- [OpenAI compatibility](https://docs.together.ai/docs/inference/openai-compatibility): Use an OpenAI SDK client against Together AI unchanged.
- [API keys and authentication](https://docs.together.ai/docs/api-keys-authentication): Create and use project-scoped API keys.
- [Serverless model catalog](https://docs.together.ai/docs/serverless/models): Models available for per-token inference.
- [Pricing](https://www.together.ai/pricing): Prices for inference, fine-tuning and GPU clusters.
## For AI agents
Every page on www.together.ai returns Markdown instead of HTML when requested with an
`Accept: text/markdown` header; there are no `.md` URLs on this domain and `.md` paths return 404.
On docs.together.ai each page has a `.md` mirror (append `.md` to the page URL). The full list of
indexable pages on this domain is at www.together.ai/sitemap.xml.
- [Full documentation text](https://docs.together.ai/llms-full.txt): The whole docs site as one Markdown file.
## Products
- [Together AI](https://www.together.ai/): Homepage: inference, fine-tuning and GPU clusters.
- [Products overview](https://www.together.ai/products): How the product lines fit together.
- [Serverless Inference](https://www.together.ai/serverless-inference): Open models on demand, priced per token.
- [Dedicated Model Inference](https://www.together.ai/dedicated-model-inference): Dedicated endpoints for a single model.
- [Dedicated Container Inference](https://www.together.ai/dedicated-container-inference): Dedicated GPUs for video, audio and avatar models.
- [Batch Inference](https://www.together.ai/batch-inference): Async large-scale jobs at up to 50% lower cost.
- [Provisioned Throughput](https://www.together.ai/provisioned-throughput): Reserved token capacity, 99% uptime SLA.
- [Fine-Tuning](https://www.together.ai/fine-tuning): Fine-tune open models, including 100B+ parameters.
- [Custom Training](https://www.together.ai/custom-training): RL and SFT post-training, full-weight or LoRA.
- [Evaluations](https://www.together.ai/evaluations): Benchmark model quality before shipping.
- [Sandbox](https://www.together.ai/sandbox): Secure code sandboxes with instant VMs.
- [GPU Clusters](https://www.together.ai/gpu-clusters): H100, H200, B200 and GB200 clusters with InfiniBand.
- [AI Factory](https://www.together.ai/ai-factory): Private GPU cloud for training, fine-tuning and inference.
- [Managed Storage](https://www.together.ai/managed-storage): Object and parallel-filesystem storage for AI.
## GPU hardware
- [NVIDIA GB200 NVL72](https://www.together.ai/gpu/nvidia-gb200-nvl72): 72 Blackwell GPUs, 13.4TB HBM3e; pricing and specs.
- [NVIDIA GB300 NVL72](https://www.together.ai/gpu/nvidia-gb300-nvl72): 72 Blackwell Ultra GPUs, 21TB HBM3e; pricing and specs.
- [NVIDIA HGX B200](https://www.together.ai/gpu/nvidia-hgx-b200): 180GB HBM3e, 7.7TB/s, FP4; on-demand or reserved.
- [NVIDIA HGX B300](https://www.together.ai/gpu/nvidia-hgx-b300): 288GB HBM3e, 8TB/s, enhanced FP4; on-demand or reserved.
- [NVIDIA HGX H200](https://www.together.ai/gpu/nvidia-h200): 141GB HBM3e, 4.8TB/s; on-demand or reserved.
- [NVIDIA HGX H100](https://www.together.ai/gpu/nvidia-h100): 80GB HBM3, 3.35TB/s; on-demand or reserved.
## Models
- [Model catalog](https://www.together.ai/models): 200+ text, image, video, code and audio models.
- [ByteDance Seedance 2.5](https://www.together.ai/models/seedance-2-5): Video model generating 30-second audio-video clips in one pass, up to 50 reference inputs.
- [DeepSeek V4 Flash 0731](https://www.together.ai/models/deepseek-v4-flash-0731): 284B-parameter (13B active) MoE, 1M context, three reasoning effort levels.
- [DeepSeek V4 Pro 0813](https://www.together.ai/models/deepseek-v4-pro-0813): 1.6T-parameter (49B active) MoE, three reasoning effort levels, speculative decoding.
- [DeepSeek V4.1 Flash 1017](https://www.together.ai/models/deepseek-v4-1-flash-1017): Updated V4.1 Flash release with faster task completion and shorter outputs, 1M context.
- [Elorian Glimpse](https://www.together.ai/models/elorian-glimpse): Visual reasoning model for images and video, tracking objects through motion and predicting what happens next.
- [GLM-5.3](https://www.together.ai/models/glm-5-3): Coding model post-trained from GLM-5.2, three thinking effort levels, 1M context.
- [GLM-5.3-Flash](https://www.together.ai/models/glm-5-3-flash): Multimodal GLM-5, 320B-parameter (18B active) MoE, hybrid attention, 1M context.
- [Kimi K3](https://www.together.ai/models/kimi-k3): 2.8T-parameter MoE, 1M context, native vision.
- [MiniMax M3](https://www.together.ai/models/minimax-m3): 1M context via sparse attention, multimodal.
- [MiniMax M3.1 Flash](https://www.together.ai/models/minimax-m3-1-flash): 428B-parameter (23B active) multimodal MoE for coding, agentic workflows and computer use, text, image and video input, 1M context.
- [Muse Spark Open](https://www.together.ai/models/muse-spark-open): Meta's coding-optimized open-weights model for agentic workflows and tool calling, multimodal input, 1M context.
- [Qwen3.8-2.4T-A95B](https://www.together.ai/models/qwen3-8-max): 2.4T-parameter MoE, 1M context, multimodal input.
- [Reflection Beam](https://www.together.ai/models/reflection-beam): 501B-parameter (23B active) MoE for software engineering and long-horizon agents, 1M context, Apache 2.0.
- [Tev1 4B Experimental](https://www.together.ai/models/tev1-4b-experimental): Experimental Qwen3.5-4B decision model that returns one letter mapped to a labeled option.
- [Boreal-H3](https://www.together.ai/models/boreal-h3): Creatify Labs' advertising video model post-trained from MiniMax H3, keyframe and multi-reference control up to 2K.
- [ByteDance Seedance 2.0](https://www.together.ai/models/seedance-2-0): Multimodal video model with synchronized audio, 4-15 second clips, multi-shot narratives.
- [DeepSeek V4 Pro](https://www.together.ai/models/deepseek-v4-pro): 1.6T-parameter (49B active) MoE, 512K context.
- [FLUX 3](https://www.together.ai/models/flux-3): Multimodal video model with synchronized audio, up to 20-second multi-shot clips.
- [GLM-5.2](https://www.together.ai/models/glm-52): 744B-parameter (40B active) MoE, 1M context.
- [GPT Image 2](https://www.together.ai/models/gpt-image-2): Image generation model with built-in reasoning, native 1K to 4K outputs.
- [Inkling](https://www.together.ai/models/inkling): 975B-parameter (41B active) multimodal MoE, 524K context.
- [MiniMax H3](https://www.together.ai/models/minimax-h3): 33B video model producing 4-15 second clips at 24 FPS, 768p default with 2K regeneration.
- [NVIDIA Nemotron 3.5 Super VL](https://www.together.ai/models/nemotron-3-5-super-vl): 120B-parameter (12B active) multimodal MoE for document, image and video reasoning with tool use.
- [Qwen3.7-Plus](https://www.together.ai/models/qwen37-plus): Multimodal agent model with 1M context, hybrid GUI and CLI agent control.
- [Qwen3.8 Flash](https://www.together.ai/models/qwen3-8-flash): Hybrid thinking model with agentic coding, vision-language understanding, 1M context.
- [Step 5 Preview](https://www.together.ai/models/step-5-preview): StepFun's 600B-parameter (27B active) open-weights MoE for long-horizon agentic work, text and vision input, 1M context.
- [Ternary Bonsai 2 27B](https://www.together.ai/models/ternary-bonsai-2-27b): Ternary Qwen3.8 27B at 1.76 bits/weight, vision input, 262K context.
- [Ternary Bonsai 27B](https://www.together.ai/models/prism-ml-ternary-bonsai-27b): Ternary Qwen3.6 27B at 1.71 bits/weight, 262K context.
## Model providers
- [Meta](https://www.together.ai/models-providers/meta): Llama 4 Maverick and Scout.
- [OpenAI](https://www.together.ai/models-providers/openai): gpt-oss-120B and gpt-oss-20B, Apache 2.0.
- [DeepSeek](https://www.together.ai/models-providers/deepseek): DeepSeek reasoning and general models.
- [Moonshot AI](https://www.together.ai/models-providers/moonshot-ai): Kimi models: long context, vision, agentic coding.
- [Z.ai](https://www.together.ai/models-providers/zai-org): GLM open-weight coding models.
- [Qwen](https://www.together.ai/models-providers/qwen): Qwen and QwQ hybrid-reasoning models.
- [MiniMax](https://www.together.ai/models-providers/minimaxai): MiniMax sparse-attention long-context models.
- [Google](https://www.together.ai/models-providers/google): Gemma open models, built from Gemini technology.
- [NVIDIA](https://www.together.ai/models-providers/nvidia): Nemotron hybrid Mamba-Transformer models.
- [Mistral](https://www.together.ai/models-providers/mistral): Mistral models, Apache 2.0, multilingual.
- [Deep Cogito](https://www.together.ai/models-providers/deep-cogito): Cogito self-improving reasoning models.
- [Alibaba](https://www.together.ai/models-providers/alibaba): Wan image and video, HappyHorse video.
- [ByteDance](https://www.together.ai/models-providers/bytedance): Seedance video and Seedream image models.
- [Black Forest Labs](https://www.together.ai/models-providers/black-forest-labs): FLUX open-weight image models.
- [Ideogram](https://www.together.ai/models-providers/ideogram): Image models with strong text rendering.
- [RunDiffusion](https://www.together.ai/models-providers/rundiffusion): Juggernaut photorealistic image models.
- [Cartesia](https://www.together.ai/models-providers/cartesia): Sonic low-latency text-to-speech.
- [Deepgram](https://www.together.ai/models-providers/deepgram): Real-time speech-to-text models.
## Solutions
- [Voice](https://www.together.ai/solutions/voice): Co-located STT, LLM and TTS for voice agents.
## Customers
- [Customer stories](https://www.together.ai/customers): Production stories on inference, tuning and clusters.
- [Cursor](https://www.together.ai/customers/cursor): Real-time inference for in-editor coding agents.
- [Decagon](https://www.together.ai/customers/decagon): Sub-second voice AI, 6x cost reduction.
- [Hedra](https://www.together.ai/customers/hedra): AI video at 60% lower cost, 3x faster.
- [Vercept](https://www.together.ai/customers/vercept): 5x computer-automation inference performance.
- [The Washington Post](https://www.together.ai/customers/washington-post): "Ask The Post AI", 1.79B tokens a month.
- [Slingshot AI](https://www.together.ai/customers/slingshot-ai): Mental-health AI fine-tuned by API, ~5x cheaper.
- [LegionEdge](https://www.together.ai/customers/legionedge): Real-time AI prototyping on Code Sandbox.
- [SCB10X](https://www.together.ai/customers/scb10x): Typhoon Thai-language models, trained on Together AI.
## Company
- [About us](https://www.together.ai/about-us): The team behind the AI Native Cloud.
- [Careers](https://www.together.ai/careers): Open roles across research and engineering.
- [News and press](https://www.together.ai/press): Press coverage and announcements.
- [Research](https://www.together.ai/research): Research on kernels, inference and agents.
- [Blog](https://www.together.ai/blog): Product launches and technical guides.
- [Events](https://www.together.ai/events): Conferences, meetups and webinars.
- [Series C announcement](https://www.together.ai/blog/announcing-our-series-c): The $800M Series C and open-model economics.
- [What is an AI Native Cloud?](https://www.together.ai/blog/what-is-an-ai-native-cloud): AI-native cloud versus legacy infrastructure.
- [Contact sales](https://www.together.ai/contact-sales): Enterprise, cluster and inference inquiries.
- [Support](https://www.together.ai/support): Support for deployment and production issues.