Models / DeepSeek

DeepSeek

Deploy DeepSeek V3 and R1 on Together AI. Transparent reasoning, frontier performance, and 90% cost savings vs closed-source alternatives.

Why DeepSeek on Together AI?

Designed for production workloads that need 
consistent performance and operational control.

Frontier performance at a fraction of the cost

DeepSeek’s MoE architecture delivers GPT-4 class performance at one-tenth the price. Together AI’s infrastructure adds 70–90% cost savings over closed-source alternatives.

Transparent reasoning, no black boxes

DeepSeek R1 exposes its complete chain-of-thought in <think> tags. Debug, verify, and trust your model’s reasoning process at every step.

Enterprise-ready from day one

SOC 2 Type II certified, HIPAA compliant, and deployed on US-based infrastructure. Full model ownership with no data retention by default.

Meet the DeepSeek family

Explore top-performing models across text, image, video, code, and voice.

New

Code

DeepSeek-V3.2-Exp

New

Chat

DeepSeek-V3.1

Chat

DeepSeek-R1-0528 Throughput

Chat

DeepSeek-V3-0324

Chat

DeepSeek-R1-0528

Chat

DeepSeek R1 Distilled Llama 70B

Chat

DeepSeek R1 Distilled Llama 70B Free

Chat

DeepSeek R1 Distilled Qwen 14B

Breakthrough technical innovations

Explore all the game-changing architectural advances that make DeepSeek models shine.

  • Mixture of Experts (MoE)

    Sparse expert routing activates only 37B out of 671B parameters for each token in V3. Advanced load balancing without auxiliary losses maintains performance while reducing computational cost.

  • Group Relative Policy Optimization

    New RL approach that removes separate value networks in RLHF, using grouped relative advantage estimation to cut compute requirements while maintaining training stability.

  • Native Reasoning Transparency

    First reasoning model to expose complete thinking process in <think> tags. Native reasoning capabilities built into model foundation through large-scale reinforcement learning.

  • MetaP Training

    First successful implementation of FP8 mixed precision training on a 671B parameter model. Pioneering reinforcement learning approach without supervised fine-tuning as preliminary step.

  • Multi-Head Latent Attention

    Innovative attention mechanism that reduces KV-cache memory requirements while maintaining modeling performance. Optimized for efficient inference deployment.

  • Multi-Token Prediction

    Novel training objective that allows the model to predict multiple tokens simultaneously. Enhanced performance and efficiency through advanced training techniques.

Deployment options

Run models using different deployment options depending on latency needs, traffic patterns, and infrastructure control.

  • Serverless

  • Inference

Serverless Inference

Real-time

A fully managed inference API that automatically scales with request volume.

Best for

Variable or unpredictable traffic

Rapid prototyping and iteration

Cost-sensitive or early-stage production workloads

Batch

Process massive workloads of up to 30 billion tokens asynchronously, at up to 50% less cost.

Best for

Classifying large datasets

Offline summarization

Synthetic data generation

Dedicated Inference

Dedicated Model Inference

An inference endpoint backed by reserved, isolated compute resources and the Together AI inference engine.

Best for

Predictable or steady traffic

Latency-sensitive applications

High-throughput production workloads

Dedicated Container Inference

Run inference with your own engine and model on fully-managed, scalable infrastructure.

Best for

Generative media models

Non-standard runtimes

Custom inference pipelines