Skip to content

Latest commit

 

History

History
247 lines (192 loc) · 6.18 KB

stable-lm-2.md

File metadata and controls

247 lines (192 loc) · 6.18 KB

Overview

Stable LM 2 is Stability AI's family of multilingual language models, released in 2024 with variants in 1.6B and 12B parameter sizes. Trained on 2 trillion tokens across seven languages, these models excel at conversational AI, tool usage, and function calling, making them ideal for RAG systems and agentic applications.

Model Variants

Stable LM 2 1.6B

  • Parameters: 1.6 billion
  • Training: Multi-stage training for improved capabilities
  • Languages: 7 (English, Spanish, German, Italian, French, Portuguese, Dutch)
  • Features:
    • Remarkably low system requirements
    • Enhanced conversational abilities
    • Tool usage and function calling
    • Updated version released 2024

Stable LM 2 12B

  • Parameters: 12.1 billion
  • Training Tokens: 2 trillion (two epochs)
  • Languages: 7 multilingual support
  • Features:
    • Medium-sized model balancing performance and efficiency
    • High performance in tool usage
    • Function calling capabilities
    • Optimized for RAG systems
    • Base and instruction-tuned variants

Supported Languages

Strong performance across seven languages:

  1. English
  2. Spanish
  3. German
  4. Italian
  5. French
  6. Portuguese
  7. Dutch

Training Details

Stable LM 2 12B

  • Training Data: 2 trillion tokens
  • Training Method: Two epochs over diverse multilingual and code datasets
  • Architecture: Decoder-only language model
  • Framework: Follows established Stable LM 2 1.6B framework

Data Composition

  • Multilingual text corpora
  • Code datasets
  • Conversational data
  • Tool usage examples
  • Function calling demonstrations

Key Capabilities

Tool Usage and Function Calling

Stable LM 2 12B Instruct:

  • High performance in tool usage scenarios
  • Function calling capabilities
  • Perfect for RAG systems as central component
  • Multi-step tool orchestration

Stable LM 2 1.6B Update:

  • Improved tool usage
  • Function calling support
  • Enhanced for agentic workflows

Conversational AI

  • Improved conversation abilities across all 7 languages
  • Natural multi-turn dialogues
  • Context retention
  • Personality consistency

Multilingual Processing

  • Balanced performance across 7 languages
  • Cross-lingual understanding
  • Code-switching support
  • Language-specific nuances

Benchmark Performance

Competitive Performance

Stability AI compared Stable LM 2 12B to:

  • Mixtral
  • Llama 2
  • Qwen 1.5
  • Gemma
  • Mistral

Results: Solid performance on zero-shot and few-shot tasks across general benchmarks outlined in the Open LLM Leaderboard.

Efficiency

  • Balances strong performance with efficiency
  • Lower memory requirements than larger models
  • Faster inference than comparable models
  • Cost-effective deployment

Use Cases

RAG Systems

  • Central model for retrieval-augmented generation
  • Document-grounded responses
  • Source citation
  • Knowledge base integration

Agentic Workflows

  • Tool orchestration
  • Function calling
  • Multi-step reasoning
  • API integration

Multilingual Applications

  • Customer support in 7 languages
  • Content generation across languages
  • Translation assistance
  • Cross-lingual information retrieval

Conversational AI

  • Chatbots and virtual assistants
  • Interactive applications
  • Customer service automation
  • Educational tutors

Deployment Options

Platforms

  • Hugging Face Hub:
    • stabilityai/stablelm-2-1_6b
    • stabilityai/stablelm-2-1_6b-chat
    • stabilityai/stablelm-2-12b
    • stabilityai/stablelm-2-12b-chat
  • Stability AI Platform: Direct access
  • Cloud providers: AWS, Azure, Google Cloud

Hardware Requirements

Stable LM 2 1.6B:

  • CPU: Modern laptops
  • GPU: Optional, improves speed
  • RAM: 4-8GB
  • Storage: ~3.2GB (FP16)
  • Remarkably low requirements

Stable LM 2 12B:

  • GPU: Recommended (consumer GPU sufficient)
  • RAM: 24-32GB
  • Storage: ~24GB (FP16)
  • Quantization: Reduces to fit smaller GPUs

Model Architecture

Design Principles

  • Decoder-only transformer
  • Optimized attention mechanisms
  • Efficient parameter allocation
  • Balanced for speed and quality

Optimization Focus

  • Memory efficiency
  • Inference speed
  • Multilingual capability
  • Tool use reliability

Integration Support

Frameworks

  • Hugging Face Transformers
  • PyTorch
  • vLLM for serving
  • LangChain (RAG, agents)
  • LlamaIndex (document retrieval)

RAG Frameworks

  • LangChain connectors
  • Vector database integrations
  • Document processing pipelines
  • Retrieval systems

Licensing

Stability AI Membership

Non-Commercial Use: Available freely

Commercial Use: Requires Stability AI Membership

Check official licensing for specific terms and conditions.

Release Timeline

  • Stable LM 2 1.6B: Initial release
  • Stable LM 2 1.6B Update: Enhanced conversational abilities, tool usage
  • Stable LM 2 12B: April 2024
  • Continuous Updates: Ongoing improvements

Comparison: 1.6B vs. 12B

Feature 1.6B 12B
Parameters 1.6B 12.1B
System Requirements Very Low Medium
Performance Good Excellent
Tool Usage Improved High Performance
Use Case Edge/Mobile Cloud/Enterprise

Performance Characteristics

Efficiency

  • Lower memory than competitors
  • Fast inference
  • Energy-efficient
  • Cost-effective operation

Quality

  • Competitive with larger models
  • Strong multilingual performance
  • Reliable tool usage
  • Good generalization

Function Calling Features

Capabilities

  • Define functions and their parameters
  • Model selects appropriate functions
  • Generate function calls with correct arguments
  • Process function results
  • Multi-step function chains

RAG Integration

  • Retrieval function calling
  • Document selection
  • Source attribution
  • Grounded generation

Research and Development

Developed by Stability AI focusing on:

  • Efficient model architectures
  • Multilingual capabilities
  • Tool use and function calling
  • Practical deployment scenarios

Pricing

Available for testing on Hugging Face. Commercial use requires Stability AI Membership - check official pricing.