sambal/qwen3.5-27b-grpo-checkpoint-440

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 5, 2026Architecture:Transformer Featherless Exclusive Cold

The sambal/qwen3.5-27b-grpo-checkpoint-440 is a 27 billion parameter Qwen3.5ForConditionalGeneration model, representing a merged Hugging Face checkpoint from global step 440 of its GRPO training. This model is provided in BF16 precision across 23 safetensors shards, including its original configuration, tokenizer, processor, and chat template. It is suitable for applications requiring a large-scale Qwen3.5 variant with a 32768 token context length, delivered without quantization or weight modifications.

Loading preview...

Qwen3.5-27B GRPO Checkpoint 440 Overview

This model is a specific checkpoint (global step 440) from the training of the Qwen3.5-27B GRPO architecture. It is provided as a full model, not a LoRA adapter, and maintains the original weight precision and configuration from its training run.

Key Characteristics

  • Architecture: Based on the Qwen3_5ForConditionalGeneration architecture.
  • Parameter Count: A substantial 27 billion parameters, indicating strong general-purpose language understanding and generation capabilities.
  • Context Length: Supports a context window of 32768 tokens, allowing for processing and generating longer texts.
  • Precision: Weights are in BF16 (bfloat16) precision, cast from an FP32-master merged checkpoint.
  • Format: Distributed as full model weights across 23 safetensors shards, ensuring comprehensive model integrity.
  • Included Components: Comes with its original configuration, tokenizer, processor configuration, and chat template, facilitating direct deployment and use.

Good For

  • Research and Development: Ideal for researchers and developers looking to experiment with a specific, unquantized checkpoint of the Qwen3.5-27B GRPO model.
  • Applications Requiring High Fidelity: Suitable for tasks where maintaining the original model's weight precision and full architecture is crucial, without the potential compromises of quantization.
  • Long Context Tasks: The 32768 token context length makes it well-suited for applications involving extensive documents, detailed conversations, or complex code analysis.

Finetunes
1 models