LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Iter16

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 29, 2026Architecture:Transformer Featherless Exclusive Cold

LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Iter16 is a 9 billion parameter Qwen3.5-based causal language model. This is an intermediate checkpoint (iteration 16) from an OPSD (Optimized Parameter-Efficient Supervised Distillation) training process, specifically for the 'trailing_user' PI. It is a fully merged model, requiring no external adapters or merge tools, and is suitable for direct loading and evaluation of its student performance.

Loading preview...

OPSD-PI Qwen3.5-9B Medium Trailing - Merged Checkpoint Iteration 16

This model, developed by LSW142857, is a 9 billion parameter Qwen3.5-based causal language model representing an intermediate checkpoint (iteration 16) from an OPSD (Optimized Parameter-Efficient Supervised Distillation) training run. Specifically, it targets the trailing_user PI (Parameter-Efficient Instruction-tuning) configuration. Unlike models requiring additional merging steps, this repository provides a fully merged checkpoint, ready for direct loading and use.

Key Characteristics

  • Fully Merged: The model shards already incorporate the merged expert-SFT initialization, the iteration-16 main-model OPSD LoRA update, the iteration-16 MTP LoRA update, and all directly trained full-MTP tensors. No external adapters or merge tools are needed.
  • Intermediate Checkpoint: This is iteration 16 of a 31-iteration training process, offering a snapshot of the model's capabilities at this stage.
  • Integrity Verified: All 775 output tensors were verified for exactness before upload, including LoRA targets, full-MTP targets, and overlapping MTP projections.

Usage and Evaluation

This model is designed for evaluating the student performance without adding the Medium PI. It is recommended to use held-out tasks for evaluation, rather than the 1024 training rows. Detailed configuration, iteration-16 metrics, and provenance are available in training_config.json and merge_manifest.json within the repository.