minjaechoi/qwen3p6-35b-a3b-2p00bit-r53

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 4, 2026Architecture:Transformer Featherless Exclusive Cold

minjaechoi/qwen3p6-35b-a3b-2p00bit-r53 is a 35.1 billion parameter language model based on the Qwen3.6-35B-A3B architecture, developed by minjaechoi. This model features an internal research checkpoint utilizing routed experts with an average of 2.00 bits, while other weights are in BF16 format. It is designed for efficient inference, loading dequantized weights in BF16 tensors with standard transformers and vLLM libraries, making it suitable for applications requiring reduced memory footprint and faster processing.

Loading preview...

Overview

minjaechoi/qwen3p6-35b-a3b-2p00bit-r53 is an internal research checkpoint derived from the Qwen/Qwen3.6-35B-A3B base model. This 35.1 billion parameter model incorporates a unique quantization strategy, where routed experts average 2.00 bits per weight, while all other weights are maintained in BF16 format. The model is designed to be compatible with standard transformers and vLLM libraries, loading its dequantized weights directly as BF16 tensors.

Key Capabilities

  • Efficient Inference: Utilizes a mixed-precision approach with 2.00-bit routed experts and BF16 weights, potentially leading to reduced memory consumption and faster inference speeds compared to full BF16 or FP16 models.
  • Standard Library Compatibility: Seamlessly integrates with popular deep learning frameworks like transformers and vLLM for easy deployment and use.
  • Research-Oriented Design: Represents an internal research checkpoint (r53) focused on exploring advanced quantization and routing techniques for large language models.

Good for

  • Memory-Constrained Environments: Ideal for deployments where memory footprint is a critical concern, benefiting from the highly quantized expert weights.
  • High-Throughput Applications: The optimized weight representation can contribute to faster processing and higher throughput in inference tasks.
  • Experimentation with Quantization: Useful for researchers and developers interested in evaluating the performance and efficiency of routed expert architectures with aggressive quantization.

Finetunes
1 models