Berkelium-ai/BerkeliumGPT2-Coder-3b-MLX

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Berkelium-ai/BerkeliumGPT2-Coder-3b-MLX is a 3.1 billion parameter Qwen 2.5 architecture model, fine-tuned by Berkelium-ai for code generation. It is trained on the BerkeliumCoding dataset, derived from 176 permissive repositories, and features a 4,096 token context window. This model is optimized for code completion and generation tasks, provided in bfloat16 full precision, and is natively compatible with both Apple MLX and Hugging Face Transformers.

Loading preview...

BerkeliumGPT2-Coder-3b-MLX: A Code-Optimized Language Model

BerkeliumGPT2-Coder-3b-MLX is a 3.1 billion parameter language model developed by Berkelium-ai, specifically fine-tuned for code generation tasks. It is based on the Qwen 2.5 architecture and utilizes bfloat16 full precision, ensuring high fidelity in its computations. The model's training involved the proprietary BerkeliumCoding dataset, which comprises code from 176 permissive repositories, making it highly specialized for programming contexts.

Key Capabilities & Features

  • Code Generation: Optimized for generating and completing code snippets, as demonstrated by its fine-tuning on a comprehensive coding dataset.
  • Architecture: Built upon the robust Qwen 2.5 model, providing a strong foundation for its performance.
  • Precision: Utilizes bfloat16 (bf16) full precision, offering a balance between performance and memory usage without quantization.
  • Context Window: Features a 4,096 token context length, allowing it to process and generate longer code sequences.
  • Platform Compatibility: Designed for native execution on both Apple MLX (leveraging Apple Silicon Unified Memory via Metal) and Hugging Face Transformers, offering flexibility for developers.
  • Format: Provided in Native SafeTensors format for ease of use and direct integration.

Good For

This model is particularly well-suited for developers and researchers working on code-related applications. Its fine-tuning on a diverse code corpus makes it an excellent choice for:

  • Code completion tools.
  • Automated code generation.
  • Assisting with programming tasks on Apple Silicon hardware due to its MLX compatibility.
  • Integration into larger code development workflows using Hugging Face Transformers.

Finetunes
1 models