i-Coder/iCoder-27B-OPSD

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

i-Coder/iCoder-27B-OPSD is a 27 billion parameter intermediate checkpoint from the iCoder-27B training pipeline, based on Qwen3.6-27B. This model is specifically designed for research into RTL design and GPU kernel optimization, representing the on-policy self-distillation (OPSD) stage. It serves as a foundational artifact for further research into its training pipeline, particularly for evaluating the impact of subsequent RLVR stages. The model is not prepared for deployment but is intended for reproducing or ablating its training stage.

Loading preview...

iCoder-27B-OPSD: An Intermediate Checkpoint for RTL and GPU Optimization Research

iCoder-27B-OPSD is a 27 billion parameter model, serving as an intermediate checkpoint within the iCoder-27B training pipeline. Developed by i-Coder, this model is specialized for tasks related to RTL design and GPU kernel optimization. It originates from the Qwen3.6-27B base model and represents the output of the on-policy self-distillation (OPSD) stage.

Key Characteristics:

  • 27 Billion Parameters: A substantial model size for complex code generation and optimization tasks.
  • Specialized Domain: Focused on hardware description languages (RTL) and GPU kernel optimization.
  • Intermediate Checkpoint: This specific version is a mid-pipeline artifact, following the SFT stage and preceding the RLVR stage in the full iCoder-27B training.

Intended Use Cases:

  • Training Pipeline Research: Primarily for researchers to reproduce, ablate, or analyze the OPSD stage of the iCoder-27B pipeline.
  • Evaluation of RLVR Impact: Useful for measuring the additional contributions of the subsequent RLVR stage.
  • No Deployment Preparation: It is crucial to note that this checkpoint is not prepared for direct deployment and is intended solely for research purposes related to its training methodology.

Finetunes
1 models