chinapao/Fawen-1.0-35B
Fawen-1.0-35B by David Zhang is a 35-billion-parameter sparse Mixture-of-Experts (MoE) language model, activating approximately 3 billion parameters per token for efficient inference. It is aggressively quantized and then retrained to recover quality, focusing on robust tool/function calling and transparent, step-by-step reasoning. This multilingual model (English and Chinese) is designed for agentic workflows and deployment-friendly applications requiring explainable AI behavior.
Loading preview...
Model Overview
Fawen-1.0-35B, developed by David Zhang, is a 35-billion-parameter sparse Mixture-of-Experts (MoE) language model. A key feature is its efficiency, activating only about 3 billion parameters per token, making its inference cost comparable to a smaller model while retaining the capacity of a larger one. The model underwent aggressive quantization followed by retraining to preserve capability, resulting in a deployment-friendly yet behavior-rich assistant.
Key Capabilities
- Tool/Function Calling: Natively trained on real-world agent interaction traces and curated tool-calling sets, enabling reliable invocation of tools for web access, file I/O, computation, and databases.
- Transparent Reasoning: Employs a built-in reasoning scaffold, separating deliberation from action. It plans using a
<think>block before executing actions via<tool_call>, making its decision-making process auditable. - Sparse Efficiency & Quantization: Achieves efficient inference and reduced memory footprint through its sparse MoE architecture and aggressive quantization-aware retraining.
- Multilingual & Long-Context: Supports both English and Chinese with an extended context window, suitable for long documents and multi-turn agent sessions.
Good For
- Assistant and agent applications requiring reliable tool use, multi-step planning, and explainable reasoning.
- Chinese/English conversational products.
- Cost-sensitive or on-premise deployments due to its optimized efficiency.