Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2

What is Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2

This is a 27B-parameter code and reasoning-enhanced model based on Qwen, optimised through multi-stage training including SFT, RLOO and MTP, focused on efficient thinking, programming and complex reasoning tasks.

Key features

• Optimised through multi-stage training including SFT, RLOO and MTP based on Qwen3.8-27B
• Significantly reduces the 94K truncation issue caused by repeated model inference
• Supports multiple quantisation formats including BF16, FP8, NVFP4, INT4, INT8 and GGUF
• Native support for text + image + video multimodal input
• Built-in MTP and DFlash2 mutually exclusive speculative decoding acceleration methods

Who it is for

Developers and researchers needing efficient programming, complex reasoning and multimodal tasks, suitable for scenarios with high requirements for model output length control and inference stability.

How to get started

1. Visit https://huggingface.co/nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2
2. Download safetensors weights in the required format (BF16, FP8, NVFP4, INT4, GGUF, etc.)
3. Load the model directly with SGLang or vLLM (FP8 version tested and working on vLLM)

Compiled from public descriptions on the tool's official website / README / launch page. Pricing and user numbers have not been verified; please check the official website. Some sections were drafted with AI assistance. Written in Chinese and English; other languages are machine-translated.