Qwen3.5-27B

The model architecture is extensively documented in official Hugging Face model cards and technical blog posts. It utilizes a novel hybrid design consisting of 64 layers organized into 16 groups, where each group contains three Gated DeltaNet (linear attention) layers and one Gated Attention layer. Detailed specifications including hidden dimensions (5120), head dimensions (128 for GDN, 256 for Gated Attention), and intermediate FFN dimensions (17408) are publicly available. While the high-level methodology of 'early-fusion' multimodal training is described, the specific pre-training recipe and architectural modifications for the vision encoder integration are less detailed than the language backbone.

Information regarding the training data remains largely high-level and lacks granular transparency. Official sources mention 'trillions of multimodal tokens' and a 'multilingual data annotation system' labeling over 30T tokens across 201 languages. However, there is no specific breakdown of dataset proportions (e.g., % web, % code, % books) or a comprehensive list of data sources. While some evaluation datasets like HLE-Verified are open-sourced, the primary pre-training corpus composition and specific filtering/cleaning methodologies are not disclosed in detail, relying on vague descriptors like 'carefully curated'.

The tokenizer is publicly accessible via the Hugging Face repository and is well-documented. It uses a Byte-level Byte Pair Encoding (BPE) approach with a vocabulary size of 248,320 (padded), significantly expanded from previous generations to support 201 languages. Vocabulary size and tokenization logic are verified through both official documentation and third-party implementations (e.g., .NET ports). The alignment between claimed language support and tokenizer efficiency is documented, though some internal token normalization details are proprietary.

As a dense model, parameter density is straightforward and clearly stated at 27.0B total parameters. Unlike the MoE variants in the Qwen 3.5 family, all parameters are active during inference. The architectural breakdown is highly detailed, specifying the exact number of layers (64), attention heads (24 Q, 4 KV), and the specific layout of Gated DeltaNet vs. Gated Attention blocks. This level of detail allows for precise calculation of computational requirements and memory overhead.

Transparency regarding training compute is extremely low. While the use of 'Next-Generation Training Infrastructure' and 'asynchronous RL frameworks' is mentioned in marketing materials, there is no public disclosure of the total GPU/TPU hours consumed, the specific hardware clusters used for the 27B variant's training, or the associated carbon footprint. The company cites efficiency gains but provides no verifiable metrics to back these claims, scoring poorly on environmental and resource transparency.

The model provides scores for standard benchmarks (MMLU-Pro: 86.1%, GPQA Diamond: 85.5%, SWE-bench Verified: 72.4%) which are verifiable through third-party leaderboards like Artificial Analysis and OpenRouter. However, the exact evaluation code, specific prompts, and few-shot examples used for official reporting are not fully public in a centralized repository. While some third-party audits exist, the lack of a comprehensive, reproducible evaluation suite from the provider prevents a higher score.

The model demonstrates high identity consistency, correctly identifying itself as Qwen 3.5 and maintaining version awareness across different deployment frameworks (vLLM, SGLang, Ollama). It is transparent about its multimodal capabilities and its position within the broader Qwen 3.5 ecosystem. There are no documented instances of the model claiming to be a competitor's product or misrepresenting its dense architecture as an MoE variant.

The model is released under the Apache 2.0 license, which is a standard, highly permissive open-source license. The license terms are clearly stated in the Hugging Face repository and official announcements, explicitly allowing for commercial use, modification, and redistribution. There are no conflicting custom terms or restrictive 'acceptable use' policies that override the base license, providing exemplary legal transparency.

Hardware requirements are well-documented by both the provider and the community. VRAM requirements for various quantization levels (Q4, Q8, FP16) are publicly available, with specific guidance for single-GPU deployment (e.g., ~16-18GB for Q4 GGUF). Memory scaling for the 262k context window is documented, and third-party tools like Unsloth provide detailed VRAM calculators. However, official documentation on the specific accuracy-performance tradeoffs of the new 'Gated DeltaNet' layers under heavy quantization is still emerging.

The model uses a clear naming convention (Qwen3.5-27B), but the changelog and version history are somewhat fragmented across blog posts and GitHub commits. While major releases are announced, minor weight updates or 'silent' optimizations (such as the March 5 GGUF update) are often communicated through third-party partners rather than a centralized, formal versioning system. There is no clear public roadmap or deprecation policy for previous versions.