Together AI

Together AI democratizes access to powerful open-source language models by providing blazing-fast inference infrastructure optimized for models like LLaMA, Mistral, and Code Llama. For teams wanting to avoid vendor lock-in or reduce API costs by 10x, Together AI is a game-changer.

The Together AI Philosophy

Together AI believes in open-source. Rather than building proprietary models, they optimize inference for the best open-source models available. This approach provides freedom: swap models, fine-tune, or self-host anytime.

Supported Models

Foundation Models

LLaMA 2 and LLaMA 3: Meta’s powerful open models

• 7B, 13B, 70B parameter versions
• Excellent general-purpose capabilities
• Strong coding abilities
• Competitive with proprietary models

Mixtral 8x7B: Sparse mixture-of-experts model

• Expert-level performance
• Efficient token generation
• Good for specialized tasks
• Open-source and available for fine-tuning

Mistral 7B: Compact but powerful

• Fastest inference
• Lowest cost
• Surprisingly capable
• Ideal for high-volume applications

Code Llama: Specialized for programming

• Code generation and explanation
• Multi-language support
• Great for developer tools
• Open weights for fine-tuning

Specialized Models

• Summarization: Aligned models for document summarization
• Legal: Fine-tuned for legal document analysis
• Medical: Specialized for healthcare applications
• Multilingual: Excellent non-English capabilities

Why Together AI Wins on Speed

Hardware Optimization

Together invested in custom inference optimization:

• Optimized for batch inference
• Reduced latency (2-5x faster than alternatives)
• Higher throughput per GPU
• Efficient memory usage

Performance Numbers

Real-World Impact

Scenario: Generate product descriptions for 100,000 items

• OpenAI: 3 hours, $500 cost
• Together AI: 45 minutes, $50 cost
• Speed improvement: 4x faster
• Cost savings: 90% reduction

Cost Efficiency

Transparent Pricing

No surprises, no hidden fees:

• LLaMA 2 7B: $0.08 per 1M input tokens
• LLaMA 2 70B: $0.70 per 1M input tokens
• Mistral 7B: $0.06 per 1M input tokens

Input tokens cost same as output (unlike some competitors).

Free Trial

• $5 free credits
• Test models before committing
• No credit card required

Volume Discounts

• 10%+ discount for $1000+/month
• Higher discounts for committed volumes
• Custom pricing for enterprise

API Integration

Easy Implementation

import together

response = together.Complete.create(
    prompt="Complete this: The future of AI is",
    model="togethercomputer/llama-2-70b-chat",
    max_tokens=256,
    temperature=0.7
)

print(response['output']['choices'][0]['text'])

Language Support

• Python (most complete)
• JavaScript/Node.js
• Java
• Go
• REST API for any language

Streaming Support

For real-time token generation (chatbots, interactive apps):

stream = together.Complete.stream(
    prompt="Write a poem about AI",
    model="togethercomputer/llama-2-70b",
)

for token in stream:
    print(token, end='', flush=True)

Fine-Tuning on Your Data

Custom Model Training

Together provides fine-tuning services:

1. Upload data: Your examples and expected outputs
2. Configure training: Set parameters and options
3. Train: Together handles infrastructure
4. Deploy: Use your fine-tuned model via API

Fine-Tuning Benefits

• Lower cost per inference (tokens cost less)
• Better performance on your task
• Faster response generation
• Domain-specific knowledge

Example: Customer Support

Fine-tune LLaMA 2 on 5,000 support conversations:

• Results: 30% accuracy improvement
• Cost: 40% less per request
• Time to market: 1 week

Real-World Use Cases

Content Generation at Scale

Publishing company needs 1000 article summaries daily:

• Use Mixtral 8x7B for summarization
• 1000 articles cost < $5/day
• OpenAI would cost $50+/day
• Annual savings: $16,500

Customer Chatbot

E-commerce platform with 1M daily queries:

• Deploy LLaMA 2 7B for first-level triage
• Together AI: $50/day
• Alternative APIs: $500-800/day
• Monthly savings: $15,000+

Code Generation

Development team using LLaMA Code for assistant:

• Faster than cloud alternatives
• Deploy models you own (no lock-in)
• Fine-tune on your codebase
• Control and privacy

Comparison: Together AI vs Alternatives

When to Use Together AI

Choose Together AI if:

• You want to avoid vendor lock-in
• Cost per token is critical
• You need fast inference
• You plan to fine-tune models
• You want to use open-source models
• You need inference at scale

Consider alternatives if:

• You need cutting-edge proprietary models
• You need vision/multimodal capabilities
• You require bleeding-edge research models
• You prioritize brand recognition

Getting Started

1. Sign up at together.ai (free account)
2. Get API key from dashboard
3. Browse models in their playground
4. Test models with free credits
5. Read documentation for your use case
6. Start building with Python/JavaScript

Advanced Features

Model Comparison Tool

Compare models side-by-side on your prompts:

• Same input, different models
• See quality and speed tradeoffs
• Find best model for your task
• Make data-driven decisions

Batch Processing

For non-urgent inference:

• 30% cost reduction
• Ideal for reports, summaries, analysis
• Scheduled execution
• Perfect for overnight processing

Conclusion

Together AI represents a fundamental shift in how enterprises think about LLM inference. Instead of being locked into one vendor’s proprietary models, you can leverage the best open-source models available. The combination of superior speed, lower costs, and complete freedom makes Together AI compelling for any organization running inference at scale. For cost-conscious teams, those needing custom fine-tuning, or companies committed to open-source, Together AI is not just an alternative—it’s the smarter choice. Start with their free trial and compare the speed and cost yourself.