aviralku/openclaw-phase1-notes-50m

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026Architecture:Transformer Featherless Exclusive Cold

The aviralku/openclaw-phase1-notes-50m is a 27 billion parameter research checkpoint derived from a local Qwen3.8 base model. It was fine-tuned using full-parameter next-token training on the balanced OpenClaw recursive-note corpus with approximately 50 million note tokens. This model is specifically optimized for note-NTP (next-token prediction) tasks, making it suitable for research into recursive note generation and processing.

Loading preview...

OpenClaw Phase-1 Notes 50M: A Research Checkpoint

The aviralku/openclaw-phase1-notes-50m is a specialized research checkpoint, a 27 billion parameter model derived from a local Qwen3.8 base. This model underwent full-parameter next-token training on the unique OpenClaw recursive-note corpus, utilizing approximately 50 million note tokens.

Key Characteristics

  • Base Model: Derived from a Qwen3.8-27B base model.
  • Training Focus: Exclusively trained for note-NTP (next-token prediction) during its Phase 1 development.
  • Training Data: Utilized a balanced OpenClaw recursive-note corpus.
  • Training Budget: Approximately 50 million note tokens were used for fine-tuning.
  • Checkpoint Details: Represents epoch 1, step 1250 of the training process.
  • Technical Specifications: Features a policy KL coefficient of 0.01 and uses bfloat16 for weight data types.
  • Export Format: Available in Hugging Face Transformers format, distributed across six safetensors shards.

Good For

  • Research in Recursive Note Generation: Ideal for researchers exploring next-token prediction within structured or recursive note-taking systems.
  • Understanding Fine-tuning on Specialized Corpora: Provides a case study for fine-tuning large language models on highly specific, domain-centric datasets like the OpenClaw recursive-note corpus.
  • Experimentation with Qwen3.8 Derivatives: Offers a starting point for further development or analysis based on the Qwen3.8 architecture with a specialized training focus.