LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Update32

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 2, 2026Architecture:Transformer Featherless Exclusive Cold

LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Update32 is a 9 billion parameter Qwen3.5-based causal language model, developed by LSW142857, resulting from 32 optimizer updates on an OPSD run. This model incorporates merged expert-SFT initialization, OPSD main-model LoRA updates, and MTP LoRA updates, with a context length of 32768 tokens. It is designed for direct loading without additional adapter or merge steps, focusing on specific training iterations and integrity. Its primary differentiator lies in its specific training methodology and detailed provenance tracking.

Loading preview...

Overview

This model, LSW142857/OPSD-PI-Qwen3.5-9B-Medium-Trailing-1024-A6000-Merged-Update32, is a 9 billion parameter Qwen3.5-based causal language model. It represents the state after 32 completed optimizer updates (training iteration 31) from a 1024-row Medium PI trailing_user OPSD run on 8×RTX A6000 GPUs. The model is provided as a fully merged Hugging Face repository, requiring no additional adapter or merge steps for deployment.

Key Characteristics

  • Architecture: Based on Qwen3.5, with 9 billion parameters.
  • Training: Underwent 32 optimizer updates (iteration 31) using an OPSD (Optimizer-Pretrained Sparse-Dense) methodology.
  • Merging: Incorporates merged expert-SFT initialization, OPSD main-model LoRA updates, and MTP (Multi-Task Pretraining) LoRA updates, along with directly trained full-MTP tensors.
  • Integrity: Detailed sha256sum checks are provided for all 775 output tensors, ensuring provenance and consistency. Configuration and finite metrics from this update are available in merge_manifest.json and training_config.json.

Usage Considerations

  • Direct Loading: Designed for direct loading using AutoModelForCausalLM and AutoProcessor from the transformers library.
  • Evaluation: The PI (Pre-training Iteration) was teacher-only during training. It is recommended to evaluate the student model without adding PI and to use held-out tasks, distinct from the 1024 training rows, for accurate assessment.

Finetunes
1 models