aqbond/deepswe-k8s-sync-S0b16-c256-replace-tokenexact-step175

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The aqbond/deepswe-k8s-sync-S0b16-c256-replace-tokenexact-step175 model is a 32 billion parameter language model, representing FSDP-merged Hugging Face weights from training step 175. This model is derived from an experiment involving same-step trajectory repair and S0b16 sync PPO, specifically designed for scenarios requiring precise token replacement. It is optimized for tasks where fine-grained control over token generation and replacement is critical, stemming from a deepSWE Kubernetes synchronization process.

Loading preview...

Model Overview

The aqbond/deepswe-k8s-sync-S0b16-c256-replace-tokenexact-step175 model is a specialized language model with 32 billion parameters. It comprises FSDP-merged Hugging Face weights captured after 175 training steps.

Key Characteristics

  • Training Origin: This model is a product of a specific experiment named deepswe-k8s-sync-S0b16-c256-replace-tokenexact-141to179-20260826-111352.
  • Training Methodology: The training recipe involved "same-step trajectory repair" using gpt-5.5 key-fork and antidrown techniques, addressing 4 failures with a k=4 factor, followed by S0b16 sync PPO.
  • Checkpoint Source: The weights were extracted from the actor component at global_step_175 within the training outputs.
  • Context Length: The model supports a context length of 32768 tokens.

Primary Use Case

This model is particularly suited for applications requiring highly precise token replacement, as indicated by its replace-tokenexact designation. Its development within a deepSWE Kubernetes synchronization framework suggests potential utility in automated code generation, refactoring, or other technical domains where exact token manipulation is crucial. It represents a specific snapshot of a complex training trajectory, offering a fine-tuned state for targeted tasks.

Finetunes
1 models