HostYourAI/loes-qwen3.8-27b
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold
HostYourAI/loes-qwen3.8-27b is a 27 billion parameter, instruction-tuned causal language model developed by HostYourAI. It is a supervised fine-tune of orcarouter/Qwen3.8-27B-Uncensored, specifically optimized for natural Dutch language and conversational AI applications. This model excels at generating human-like responses in Dutch, making it suitable for research and evaluation in Dutch and English conversational tasks.
Loading preview...
Loes: A Dutch Conversational AI by HostYourAI
Loes is a 27 billion parameter AI assistant developed by HostYourAI, specifically fine-tuned for natural Dutch language and conversational interactions. Hosted on EU soil, this model is a supervised fine-tune of orcarouter/Qwen3.8-27B-Uncensored.
Key Capabilities & Features
- Dutch Language Optimization: Primarily focused on generating natural and conversational Dutch responses.
- Multimodal Base: Inherits multimodal capabilities (vision tower) from the base Qwen3.8-27B model, though only language components were fine-tuned in this release.
- Instruction-Tuned: Trained on
HostYourAI/loes-xl-52k, a dataset comprising 51,717 Dutch instruction pairs, including synthetic conversations, Aya Collection examples, and xP3x tasks. - Full Model Merge: The repository contains the fully merged model in bfloat16, allowing direct loading without separate LoRA adapters.
Intended Use Cases & Limitations
- Research and Evaluation: Ideal for research and evaluation in Dutch and English conversational AI applications.
- Conversational AI: Suited for tasks requiring natural language generation and understanding in Dutch.
- Uncensored Base: Users should be aware that the upstream checkpoint is uncensored, and the model may not reliably refuse unsafe or harmful requests; application-specific safety measures are recommended.
- Context Length: While the inherited architecture supports longer contexts, this fine-tune was trained with a maximum length of 1,024 tokens and has not been systematically evaluated for longer contexts.
- Potential for Hallucination: Like other LLMs, it may hallucinate, reproduce biases, or provide incorrect information.