Cod Strol Mamba: New AI Coding Model

Jul 18, 2024

Cod Strol Mamba: New AI Coding Model

Introduction

  • Cod Strol Mamba: New large language model for coding.
  • Parameters: Over 7 billion.
  • Architecture: Based on Mamu architecture.
  • License: Apache 2.0 (commercial use allowed).
  • Token Context Window: 256k tokens, larger than previous models.

Performance & Features

  • Inference: Faster inference for larger context tasks.
  • Comparison: Smaller models (7 billion parameters) offer faster speeds and lower costs.
  • Benchmark: Scores 75% in human evaluation benchmark vs GPT-4 Omni's 90%.
  • Additional Model: Mastl 7 billion parameter model, best open-source math model.
  • Performance Range: Cod Strol Mamba has top-tier performance among similar-sized models.

Access & Usage

  • Preview: Access via Mistal AI console under the Cod tab.
  • Verification: Requires phone number verification.
  • API Keys: Obtainable after signing up and verification.
  • Chat Interface: Cod Strol Mamba will be integrated shortly.
  • Installation: Locally via platforms like LM Studio.
  • Quantized Models: Availability expected within 24 hours.
  • Steps to Install:
    • Install LM Studio.
    • Search for Cod Strol Mamba model.
    • Download and use in the chat tab.

Model Characteristics

  • New Family: Represents a new architectural direction for Mistal AI.
  • Free Use: Modifiable and distributable.
  • Efficiency: Faster for extensive user engagement.
  • Training: Advanced code and reasoning capabilities.

Performance Metrics

  • Outpaces models like Code Gamma, Code Lama 7B, Deep Seek 1.5 7B.
  • Competitive against larger models like Code Lama 34B.
  • Context Capabilities: Effective with up to 256k tokens.

Deployment & SDK

  • Inference SDK: Provided by Mistal AI, available on GitHub.
  • Platform Support: Includes Nvidia’s Tensor RT and local inference.
  • Llama CPP: Support available right away.
  • Raw Weights: Downloadable from Hugging Face.

Wrap-up and Additional Info

  • Impact: Significant advancement for coding-based models.
  • Future Videos: More content will be created to test the model.
  • Social Links:
    • Patreon: For subscriptions and more content.
    • Twitter: Daily AI news updates.
  • Call to Action:
    • Subscribe, turn on notifications, like the video.
    • Stay updated with AI news.

  • Closing: Positivity, appreciation, and more content announcements.