MOSS
Free11.8k
FreeFree tier
Inputs: textOutputs: text, image
About MOSS
MOSS is an open-source, tool-augmented conversational language model developed by Fudan University. It is designed to handle multi-turn dialogue, follow instructions, and avoid harmful requests. The model is available in several variants, including base, supervised fine-tuned (SFT), and plugin-enhanced versions that can utilize external tools such as search engines, text-to-image generators, calculators, and equation solvers. Quantized versions (4-bit and 8-bit) are also provided to reduce memory requirements for inference. The project includes model weights, training data, and engineering solutions for local deployment and fine-tuning.
Key Features
Open-source conversational language model
Tool-augmented capabilities (search engine, text-to-image, calculator, equation solver)
Multi-turn dialogue and instruction following
Available in multiple model variants (base, SFT, plugin, quantized)
Supports local deployment with quantized versions for reduced GPU memory usage
Includes preference model for improved factuality and safety (to be released)
Pros & Cons
Pros
- Fully open-source with model weights, data, and code available
- Tool-augmented capabilities extend beyond pure text generation
- Quantized versions enable inference on consumer GPUs (e.g., 12GB VRAM for 4-bit)
- Active development from an academic institution (Fudan University)
- Includes preference-based training for improved output quality
Cons
- Requires significant computational resources for full model inference
- Free tier limits and usage policies should be verified as the project evolves
- Output quality may vary depending on the specific model variant and prompt
- Plugin capabilities depend on external services (e.g., search engine, image generator) that may have their own limitations
- Documentation and community support may be less extensive than commercial alternatives
Best For
Building conversational AI assistantsDeveloping tool-using agents that can search the web or generate imagesResearch in open-source language model fine-tuning and deploymentEducational projects exploring large language model capabilitiesCreating custom chatbots with instruction-following behavior
FAQ
What hardware is needed to run MOSS locally?
Based on available information, the 4-bit quantized versions require approximately 12GB of GPU memory for inference, while 8-bit versions require about 24GB. Full precision models will need more. Exact requirements should be verified in the project documentation.
Is MOSS completely free to use?
The model weights, code, and data are released under open-source licenses (see LICENSE and MODEL_LICENSE files). There is no indication of paid tiers, but users should review the licenses for any usage restrictions.
What plugins does MOSS support?
The plugin-enhanced variant (moss-moon-003-sft-plugin) supports using a search engine, text-to-image generation, a calculator, and an equation solver, according to the project description.
Can I fine-tune MOSS on my own data?
Yes, the repository includes a fine-tuning script (finetune_moss.py) and training data (SFT_data). Users should check the provided licenses and data license for terms.
Does MOSS support languages other than Chinese?
The model is trained on high-quality Chinese and English corpora, so it should handle both languages. The exact language coverage should be verified in the documentation.
How does MOSS compare to other open-source conversational models?
MOSS is notable for its tool-augmented capabilities and academic origin. Performance comparisons depend on the specific task and model variant. Users should evaluate based on their own requirements.