yasserrmd/glm5.1-distill

Overview

yasserrmd/glm5.1-distill is a 1.2 billion parameter instruction-tuned chat model, independently fine-tuned by Mohamed Yasser. It leverages the efficient LFM2.5-1.2B-Base architecture from LiquidAI and is trained on a 50k subset of the GLM-5.1-Reasoning-1M-Cleaned dataset, which contains reasoning-style chat data distilled from larger GLM-5.1 models. The primary goal of this distillation is to enable conversational reasoning behavior in a compact model that can run on consumer GPUs, edge devices, or via quantized runtimes.

Key Capabilities

• Lightweight Reasoning: Designed for general assistant-style chat with a focus on step-by-step answers and explanations.
• Efficient Architecture: Built on the LFM2 (hybrid conv + attention) architecture, making it suitable for resource-constrained environments.
• Instruction-Tuned: Supervised-fine-tuned (SFT) to follow instructions and engage in chat-based interactions.
• Flexible Deployment: Supports various deployment methods including ONNX, GGUF, or MLX for optimized inference.

Intended Use Cases

• General Assistant Chat: Ideal for basic conversational tasks and answering questions.
• On-Device/Edge Deployment: Excellent choice for applications where a small, efficient 1.2B parameter model is necessary.
• Further Fine-tuning: Can serve as a strong base checkpoint for domain-specific fine-tuning.

Limitations

• Inherits biases and limitations from its base model and training data.
• Performance on complex reasoning, long-context tasks, or code generation will be weaker compared to larger models.
• Primarily English-centric; performance in other languages may vary.
• Not safety-aligned or production-ready; can confidently hallucinate facts.