Qwen3.6-27B-FP8

Qwen3.6-27B-FP8 is the official fine-grained FP8-quantized release of Alibaba's Qwen3.6-27B dense model, published on Hugging Face under the Apache 2.0 license in April 2026. It uses block-size-128 FP8 (F8_E4M3) quantization that Alibaba reports preserves performance nearly identical to the BF16 reference, with independent studies showing mean KL divergence of ~5.5e-3 bits per token versus BF16. The underlying model is a 27-billion-parameter dense multimodal architecture combining a vision encoder with a hybrid language backbone: 64 layers arranged as 16 blocks of Gated DeltaNet linear attention alternating with Gated full attention and FFN, plus a trained Multi-Token Prediction (MTP) head for speculative decoding.

On official benchmarks, Qwen3.6-27B delivers flagship-level agentic coding for its size: SWE-bench Verified 77.2 (beating the prior 397B MoE flagship at 76.2), Terminal-Bench 2.0 59.3 (matching Claude 4.5 Opus), LiveCodeBench v6 83.9, MMLU-Pro 86.2, GPQA Diamond 87.8, and AIME 2026 94.1. It natively supports 262,144-token context (extensible to ~1M via YaRN) and multimodal inputs including text, images, and video, scoring 82.9 on MMMU and 87.7 on VideoMME. Community consensus positions it as one of the strongest open-weight dense models of early 2026, especially for local deployment where its FP8 variant balances quality and hardware efficiency.

The FP8 checkpoint is optimized for production inference through vLLM, SGLang, and Transformers, with official guidance to enable MTP speculative decoding for higher throughput. Independent Blackwell GPU benchmarks report peak aggregate throughput exceeding 2,000 tokens/sec at high concurrency with FP8+MTP, making it a practical choice for agentic coding, reasoning, and multimodal workloads on accessible hardware compared to larger MoE alternatives.