NotHereNorThere/Coral-v1.5-0.6B-raw

TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 4, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

NotHereNorThere/Coral-v1.5-0.6B-raw is an 0.8 billion parameter experimental language model based on the Qwen3-0.6B architecture, created by NotHereNorThere. This raw TIES merge of five Qwen3-0.6B finetunes is provided before any post-merge fine-tuning or 'healing'. It showcases the unrefined state of a fresh merge, exhibiting active but unfocused reasoning and occasional identity hallucination. This model is not recommended for direct use but serves as a transparent look into the merge process.

Loading preview...

Coral-v1.5-0.6B-raw: An Experimental Pre-Heal Merge

Coral-v1.5-0.6B-raw is an experimental 0.8 billion parameter model from NotHereNorThere, built upon the Qwen3-0.6B architecture. It represents the raw, unhealed output of a TIES merge of five distinct Qwen3-0.6B finetunes. This version is provided to demonstrate the characteristics of a fresh merge before any post-merge fine-tuning or 'healing' processes have been applied.

Key Characteristics (Pre-Heal)

As an unrefined merge, this model exhibits several notable behaviors:

  • Reasoning Style: <think> blocks are active but lack focus, indicating a nascent reasoning structure.
  • Identity Hallucination: The model may confidently invent its own origin or identity.
  • Counting Errors: Lists often contain incorrect numbering or unclosed formatting.
  • Context Retention: Ability to maintain context across multiple turns is unreliable.
  • Arithmetic: Attempts at arithmetic show structural understanding but frequently produce incorrect digits.
  • De-alignment: Donor models have removed refusals, resulting in uncensored output.

Merge Recipe

The model was created using the TIES merge method on a Qwen/Qwen3-0.6B base, utilizing mergekit. The merge incorporated five donor models, each contributing to specific aspects:

  • Thinking/Reasoning: reaperdoesntknow/Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT
  • Claude-style CoT: MihaiPopa-1/Qwen-3-0.6B-Claude-4.7-Opus-Distilled
  • Code: suayptalha/Qwen3-0.6B-Code-Expert
  • De-alignment: DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored and huihui-ai/Huihui-Qwen3-0.6B-abliterated-v2

Intended Use

This raw version is not recommended for direct use in applications due to its unrefined nature and known limitations. It is primarily intended for developers and researchers interested in understanding the intermediate state of a TIES merge before fine-tuning. For a production-ready version with these issues addressed, refer to Coral-v1.5-0.6B.

Finetunes
1 models