AI Briefing
KO

Coding-specialized model Code Llama released

·2023.08.25 09:00

Key point

Code Llama, an open-source model family based on Llama 2 and optimized for coding tasks, has been released.

Details

Code Llama is a family of models with 7B, 13B, 34B parameter scales, based on Llama 2 and specialized for coding tasks.

Key Features:

  • Training Data: Trained using approximately 500 billion code tokens, including a Python-specialized model and an Instruction model that follows natural language instructions.
  • Supported Languages: Shows excellent performance in Python, C++, Java, PHP, C#, TypeScript, Bash, and more.
  • Context Extension: Supports a default 16k context window, expandable up to 100,000 tokens through RoPE scaling technology.
  • Ecosystem Integration: Immediately compatible with Hugging Face's transformers, Text Generation Inference (TGI), VS Code extensions, and more.
huggingfaceopen-sourcecode-llmllama-2codellama
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditHemmingway-1, a 27B Creative Writing Model Based on Qwen3.8, ReleasedHuggingFace BlogNeoMME Releases Single EncoderHuggingFace BlogTraining a Coding Model to Paint Watercolors Using TRL and OpenEnvRedditOpen-Weight CLM Released as Jev Alternative, Agent Decision-Making 4x FasterPyTorchKRTrillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative DecodingHacker NewsJevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection