LSW142857/OPSD-PI-Qwen3.5-9B-Strong-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-Strong-Trailing-1024-A6000-Merged-Update32 is a 9 billion parameter Qwen3.5-based causal language model. This model represents the 32nd optimizer update from an OPSD (Optimized Parameter Space Dilation) training run, specifically utilizing a Strong PI (Personalized Instruction) trailing user configuration. It is provided as a fully merged Hugging Face model, requiring no additional adapter or merge steps, and is suitable for direct inference in applications requiring a robust Qwen3.5 variant.

Loading preview...

Model Overview

This model, LSW142857/OPSD-PI-Qwen3.5-9B-Strong-Trailing-1024-A6000-Merged-Update32, is a 9 billion parameter Qwen3.5-based causal language model. It is the result of 32 completed optimizer updates (training iteration 31) from an OPSD (Optimized Parameter Space Dilation) training run. The training utilized a "Strong PI trailing user" configuration on 8x RTX A6000 GPUs, processing 1024 rows of data.

Key Characteristics

  • Fully Merged: The repository contains a directly loadable Hugging Face model, with all components (expert-SFT initialization, OPSD main-model LoRA update, MTP LoRA update, and full-MTP tensors) merged into a single model. No additional adapter or merge steps are required for deployment.
  • Provenance and Integrity: The model's integrity can be verified using sha256sum against the provided SHA256SUMS file. Detailed configuration, source identity, and finite metrics from this update are available in merge_manifest.json and training_config.json.
  • Teacher-Only PI Training: During its training, the Personalized Instruction (PI) was teacher-only. This implies that the model should be evaluated on held-out tasks, separate from the 1024 training rows, and without adding PI during evaluation.

Good For

  • Developers seeking a pre-merged, ready-to-use Qwen3.5-9B variant from a specific OPSD training run.
  • Research and development involving models trained with Personalized Instruction (PI) and Optimized Parameter Space Dilation (OPSD) methodologies.
  • Applications where a 9B parameter model with a 32k context length is suitable for causal language modeling tasks.

Finetunes
1 models