NotHereNorThere/Coral-v1.6-0.6B

Coral-v1.6-0.6B Overview

Coral-v1.6-0.6B is a 0.8 billion parameter model developed by NotHereNorThere, building upon its predecessor, Coral-v1.5-0.6B. Unlike previous iterations, v1.6 is a pure fine-tune experiment, not a new merge, focusing on enhancing reasoning capabilities through 2,000 rows of multi-domain reasoning data. The base architecture is derived from a TIES merge of Qwen3 finetunes.

Key Capabilities & Differentiators

• Reinforced Chain-of-Thought (CoT): The fine-tuning process successfully reinforced CoT, making it consistent and reliable for structured work, even if it's "always-on" and verbose.
• Structured Reasoning: Excels at multi-step logical deduction, as demonstrated by its performance on puzzles like "Bat and ball" and "Bloops / razzles transitivity."
• Uncensored Behavior: Designed to answer edge content confidently without refusal, though factual accuracy on such content may vary.
• Training Data Focus: Utilizes reasoning traces from DeepSeek-R1 and Gemini 2 Flash Thinking (via QuixiAI/dolphin-r1) to teach the shape of good reasoning rather than distilling frontier model knowledge, which is less effective at this scale.

Good For

• Reasoning Tasks: Ideal for applications requiring consistent, structured reasoning and multi-step logic, particularly where the model's thought process (CoT) is beneficial.
• Small-Scale Deployment: Its 0.8B parameter size makes it efficient for deployment on resource-constrained hardware, with Q6_K quantization showing near-reference quality.
• Experimental Use: Serves as a stepping stone for future Coral models, offering insights into the impact of pure reasoning-focused fine-tuning on smaller models.