Thalia-70B-0307-Clean is a 70 billion parameter chat model developed by Nabbers1999, merged from Lumimaid-v0.2-70B, Strawberrylemonade-L3-70B-v1.1, and DeepSeek-R1-Distill-Llama-70B. Built on an unsloth/Llama-3.3-70B-Instruct base, it combines the creative writing abilities of its chat model parents with the deep reasoning capabilities of Deepseek. This model is designed to function as a hybrid thinker, excelling in tasks requiring both creative generation and logical thought, with a notable emphasis on prefilling the tag for optimal performance.
Loading preview...
Thalia-70B-0307-Clean: A Hybrid Reasoning and Creative Chat Model
Thalia-70B-0307-Clean is a 70 billion parameter chat model developed by Nabbers1999, created as a distillation model for future projects. It is a merge of several pre-trained language models, combining their strengths to offer a unique set of capabilities.
Key Capabilities
- Hybrid Thinking: Merges the creative writing prowess of Lumimaid and Strawberrylemonade with the deep reasoning abilities of DeepSeek-R1-Distill-Llama.
- Enhanced Reasoning: Specifically designed to improve logical thought processes, building upon its Deepseek ancestry.
- Creative Writing: Retains and enhances the creative generation aspects from its chat model parents.
- Safety Alignment: The merging process has addressed and "healed" previous safety alignments, aiming for more robust and reliable outputs.
Merge Details
This model was created using the mergekit tool, employing the DARE TIES merge method. The base model for this merge was unsloth/Llama-3.3-70B-Instruct. The primary components merged include:
- NeverSleep/Lumimaid-v0.2-70B
- deepseek-ai/DeepSeek-R1-Distill-Llama-70B
- sophosympatheia/Strawberrylemonade-L3-70B-v1.1
Usage Recommendation
For optimal performance, users should prefill the opening <think> tag when interacting with Thalia-70B-0307-Clean. This is crucial because its hybrid ancestry means it may sometimes initiate thought processes without explicit tagging, and prefilling ensures consistent and intended behavior.