Qwen3.5 27B

On February 25, 2026, Alibaba's Qwen team released the Qwen3.5 Medium Model series, a set of four LLMs with agentic tool-calling support. Qwen3.5-27B was named as one of the four models and is available for download on Hugging Face and ModelScope under the Apache 2.0 license for commercial use by enterprises and independent developers. The Qwen3.5 Medium series is positioned to compete with Western proprietary models, reportedly beating OpenAI's GPT-5-mini and Anthropic's Claude Sonnet 4.5 on third-party benchmark tests at similarly sized parameter counts. Qwen3.5-27B is part of a family using a hybrid Gated Delta Network plus sparse Mixture-of-Experts architecture. The series emphasizes near-lossless accuracy under 4-bit weight and KV cache quantization, with the flagship Qwen3.5-35B-A3B capable of exceeding 1 million token context on consumer GPUs with 32GB of VRAM. A fourth variant, Qwen3.5-Flash, is proprietary and accessible only via the Alibaba Cloud Model Studio API, though priced below comparable Western models.

Alibaba's Qwen (Tongyi Lab) announced four Qwen3.5 medium models on February 25, 2026: Qwen3.5-122B-A10B, Qwen3.5-27B, Qwen3.5-35B-A3B, and Qwen3.5-Flash, as confirmed by posts from the official @Alibaba_Qwen and @Alibaba_TongyiLab accounts. The Qwen3.5 series originally launched on February 16, 2026 with Qwen3.5-397B-A17B as its first open release, and the Medium series expands the open-source family. The announcement positions Qwen3.5-35B-A3B as surpassing the prior Qwen3-235B-A22B-2507 and Qwen3-VL-235B-A22B, framing the release as "more intelligence, less compute." The Medium series highlights architectural efficiency, with Qwen3.5-35B-A3B using only a fraction of its total parameters active per token thanks to its MoE design. The open-source models are distributed on Hugging Face and ModelScope under Apache 2.0, enabling commercial deployment. The accompanying Qwen3.5-Flash variant is gated behind the Alibaba Cloud Model Studio API rather than released as open weights.

Qwen3.5-27B is available as an open-weights model on Ollama's library with 21.2 million downloads, 27.8 billion parameters, and a Q4_K_M quantization packaged at 17GB under Apache License 2.0. The page is tagged with the qwen35 architecture and serves the upstream Qwen3.5 model card describing unified vision-language foundation training and an efficient hybrid architecture. The bundled model card details Qwen3.5 enhancements including early-fusion multimodal training, gated Delta Networks combined with sparse mixture-of-experts for throughput, and reinforcement learning scaled across million-agent environments. It also reports 201 supported languages and a benchmark table that positions Qwen3.5-397B-A17B against flagship peers on MMLU-Pro, IFEval, and related evaluations, giving developers concrete capability references for the 27B sibling.