Qwen Qwen3.5 Small Series

Overview

Qwen3.5 Small Series is Alibaba's Qwen team's March 2, 2026 release of compact language models spanning 0.8B to 9B parameters, designed explicitly for on-device deployment, edge computing, and privacy-sensitive environments. Building on the flagship Qwen3.5-397B architecture released in February 2026, this small series distills the same unified vision-language foundation and hybrid attention design into a range of model sizes that run on consumer-grade GPUs, mobile chips, and edge devices. The release reflects a deliberate shift away from the industry's parameter-scaling arms race, prioritizing architectural efficiency and advanced training techniques like Scaled Reinforcement Learning.

The four models (Qwen3.5-0.8B, 2B, 4B, and 9B) are available as both Base and Instruct variants on Hugging Face Hub and ModelScope under the Apache 2.0 license.

What's New

Scaled Reinforcement Learning in the 9B Flagship

Qwen3.5-9B is the headline model of this release, trained with Scaled Reinforcement Learning (RL) rather than conventional supervised fine-tuning alone. Where standard SFT teaches the model to mimic high-quality text, Scaled RL uses reward signals to optimize for correct reasoning paths. The practical outcomes are measurable: benchmark results on MMLU-Pro (82.5), GPQA Diamond (81.7), and IFEval (91.5) place Qwen3.5-9B above much larger competing models—including GPT-OSS-120B on several metrics—while generating tokens faster due to its compact size. The official model card highlights Scalable RL Generalization as a key training advancement; complex multi-step instruction following is significantly more reliable, reflected in the 91.5 IFEval score.

Unified Vision-Language Foundation (Early Fusion)

All four Qwen3.5 Small models include a Vision Encoder and are classified as image-text-to-text models on Hugging Face, matching the multimodal architecture of the flagship Qwen3.5-397B. Unlike earlier adapter-based approaches that bolted a separate vision encoder (like CLIP) onto a language backbone, Qwen3.5 processes text and visual tokens within the same latent space from the earliest training stages. This early-fusion design yields measurable gains: Qwen3.5-9B scores 89.2 on OCRBench, 55.5 on ERQA (spatial reasoning), and 80.3 on RealWorldQA. Vision capability scales with model size—the 4B and 9B variants are particularly well-suited for agentic workflows involving UI navigation, document analysis, and screenshot interpretation.

Ultra-Compact Edge Models (0.8B and 2B)

Qwen3.5-0.8B and Qwen3.5-2B are optimized specifically for high-throughput, low-latency scenarios on hardware with tight VRAM budgets. Through refined dense token training, these models achieve a reduced memory footprint compatible with mobile SoCs and edge devices. Both include a Vision Encoder per their official model cards; for purely text-based workloads, running text-only inference eliminates KV cache overhead from vision tokens. These compact variants deliver capable instruction following and reasoning for their scale, making them practical for always-on assistant features and embedded AI applications.

Gated DeltaNet + Sparse MoE Hybrid Architecture

All four small models inherit the Gated DeltaNet combined with sparse Mixture-of-Experts (MoE) hybrid architecture introduced in Qwen3.5-397B. While the small models use dense configurations rather than MoE at this scale, the Gated DeltaNet attention mechanism contributes to faster inference by enabling linear attention paths for long contexts. The 9B model natively supports a context length of 262,144 tokens.

Context Length Up to 262K Tokens

Qwen3.5-9B natively supports up to 262,144 tokens of context, with RoPE scaling (YaRN) enabling extension to approximately 1,010,000 tokens for long-horizon tasks. This is exceptional for a 9B-class model and allows the small series flagship to handle extended codebases, long documents, and multi-session conversations that would otherwise require much larger models.

Availability & Access

Qwen3.5 Small Series models are available immediately as open weights on Hugging Face and ModelScope. Both Base and Instruct variants are released for all four sizes. API access is available via Alibaba Cloud Model Studio (DashScope) using an OpenAI-compatible endpoint. The models can also be served locally using vLLM, SGLang, KTransformers, or Hugging Face Transformers.

System Requirements & Limitations

• Qwen3.5-9B: Weight loading at BF16 precision requires roughly 18–20 GB VRAM; quantized (GGUF/AWQ) variants reduce this to 4–8 GB. Running at the full 262K native context length adds significant KV cache overhead—reduce max_model_len or use quantization to manage memory in practice.
• Qwen3.5-4B: Weight loading requires approximately 8–10 GB VRAM at full precision; quantized variants can run on 4 GB hardware.
• Qwen3.5-2B / 0.8B: Low VRAM footprint suitable for edge devices; exact requirements vary with context length and quantization level.
• Thinking mode by model size: 0.8B and 2B default to non-thinking mode (direct response); enable thinking via enable_thinking: True. 4B and 9B default to thinking mode; disable via enable_thinking: False in the API call.
• Qwen3.5 does not officially support the /think and /nothink soft switches used in Qwen3.

Pricing & Plans

All Qwen3.5 Small Series models are released as open weights under the Apache 2.0 license, meaning the model weights are free to download and use for commercial and research purposes without royalty payments.

Self-hosted: Free. Download weights from Hugging Face or ModelScope. Inference costs are limited to your own compute.

API via Alibaba Cloud Model Studio (DashScope): Follows standard DashScope token-based pricing. Specific per-token rates for the small models are set by Alibaba Cloud and subject to change; check the DashScope pricing page for current rates.

Qwen Chat (chat.qwen.ai): Free web access for conversational use. Rate limits may apply depending on account tier.

Best For

• Developers building privacy-first AI assistants that process sensitive data locally without sending it to external APIs
• Mobile and IoT engineers deploying on-device AI features with strict VRAM and power consumption budgets
• AI agent developers who need multimodal understanding (UI navigation, document analysis) in a compact, locally runnable model
• Researchers and fine-tuners seeking an open-weight base model with frontier-class reasoning performance at an accessible scale
• Enterprises evaluating cost-efficient self-hosted LLM alternatives to large cloud-only models

FAQ