groxaxo/MagiSeek-Pro-V1

TEXT GENERATIONConcurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

groxaxo/MagiSeek-Pro-V1 is a 24 billion parameter language model built on the Mistral architecture, featuring a native context window of up to 131,072 tokens. Developed by groxaxo through a 5-phase chained QLoRA curriculum, it integrates creative writing capabilities from Magidonia with DeepSeek-style agentic reasoning. This bf16 merged checkpoint excels at general instruction following, creative writing, and complex agentic/tool-use tasks requiring long-context understanding.

Loading preview...

Overview

groxaxo/MagiSeek-Pro-V1 is a 24 billion parameter model based on the Mistral architecture, offering a substantial native context length of 131,072 tokens. It was developed through a unique 5-phase chained QLoRA curriculum, starting from WarlordHermes/Magidonia-24B-v4.3-creative-ORPO. Each phase built upon the previous, ensuring knowledge retention and progressive skill acquisition.

Key Capabilities

  • Advanced Reasoning: Incorporates DeepSeek-style step-by-step reasoning, particularly for agentic tasks.
  • Tool Use: Learned to effectively call tools through dedicated training on real tool-execution transcripts and distilled Claude Opus 4.6 reasoning traces.
  • Creative Writing: Retains strong creative writing abilities inherited from its Magidonia base.
  • Long Context: Supports up to 131,072 tokens, enabling complex interactions and detailed analysis over extended inputs.
  • Full Precision: Released as a bf16 merged checkpoint, providing a clean, full-precision snapshot suitable for downstream quantization.

Performance

The model demonstrates solid performance across various benchmarks, including an overall MMLU score of 0.7433, with strong results in MMLU-Social Sciences (0.8400) and PIQA (acc_norm 0.8413).

Intended Use

MagiSeek-Pro-V1 is designed for general instruction following, creative writing, and agentic/tool-use tasks that benefit from long-context reasoning. It is recommended as a source for further quantization (GGUF/GPTQ/AWQ) rather than direct low-VRAM deployment due to its bf16 precision and size.

Finetunes
1 models