Qwen3.5-0.8B

The model's architecture is extensively documented in the official Hugging Face repository and technical blog. It utilizes a sophisticated hybrid design (Gated Delta Networks and Gated Attention) with a specific 6×(3×DeltaNet→FFN→1×Attention→FFN) pattern. Key hyperparameters such as hidden dimensions (1024), layer count (24), and head dimensions are explicitly stated. It also discloses the use of Multi-Token Prediction (MTP) during training, which is a significant technical detail often omitted by competitors.

While the provider mentions training on a 'significantly larger scale' of multimodal tokens with 'stricter filtering' and support for 201 languages, the specific dataset composition (e.g., exact percentages of web, code, or vision data) is not disclosed. There is no public list of data sources or a detailed breakdown of the training mixture, falling into the 'general categories mentioned' tier of the scoring rubric.

The tokenizer is publicly available on Hugging Face (tokenizer.json) and is well-documented. It uses a Byte-level BPE approach with a specific vocabulary size of 151,669 tokens (padded to 248,320). The documentation clearly explains the handling of control tokens (like <|im_start|>) and its efficiency across the 201 supported languages. Third-party implementations (e.g., KerasHub, .NET) further verify its integrity.

The model clearly states its 0.8B parameter count. Unlike the larger MoE variants in the Qwen 3.5 family, this variant is dense, which is explicitly clarified in technical discussions. Detailed architectural breakdowns (KV heads, attention vs. linear layers) are provided, allowing for a clear understanding of parameter distribution, though a precise weight-by-weight breakdown is not in the primary model card.

There is a near-total lack of transparency regarding the specific compute resources used for the 0.8B variant. While the 'Next-Generation Training Infrastructure' is mentioned as a marketing highlight, there are no disclosures regarding total GPU hours, hardware counts, energy consumption, or carbon footprint. This information is conspicuously absent from the official technical report and model cards.

The model provides detailed scores across a wide array of benchmarks (MMLU-Pro, GPQA, Video-MME) and specifies the 'Thinking' vs 'Non-thinking' modes for each. However, while some evaluation settings (top_p, temperature) are disclosed, the full evaluation code and exact prompt templates for all benchmarks are not centrally hosted in a reproducible repository, requiring users to rely on third-party frameworks like OpenCompass for verification.

The model exhibits high identity consistency, correctly identifying itself as part of the Qwen 3.5 family. It maintains clear versioning and distinguishes between its base and chat variants. Documentation and system prompts (where applicable) reinforce its identity as a multimodal model from Alibaba Cloud without attempting to mimic competitors.

The model is released under the Apache 2.0 license, which is explicitly stated and included in the Hugging Face repository. This is a standard, highly permissive open-source license with no hidden 'custom' restrictions or conflicting terms, providing maximum clarity for both commercial and research use.

Hardware requirements are exceptionally well-documented by both the provider and the community (e.g., Unsloth, Ollama). VRAM requirements for various quantization levels (FP16, Q8, Q4) and context lengths (up to 262k) are publicly available. The documentation also addresses the memory scaling impact of its hybrid architecture, which is critical for a model of this size.

The model uses clear semantic versioning (Qwen3.5-0.8B), and the Hugging Face commit history provides a basic changelog. However, there is no formal, centralized changelog detailing behavioral drift or specific performance changes between minor weight updates, making it difficult for downstream users to track subtle shifts in model behavior over time.