Qwen3.5 0.8B

Qwen3.5-0.8B is a vision-language model (VLM) that integrates a causal language model with a vision encoder, designed for multimodal understanding and reasoning tasks. It is the smallest variant in the Qwen3.5 family, released on 2026-03-02, and is built on a unified vision-language foundation with early fusion training on multimodal tokens, achieving cross-generational parity with prior Qwen3 models across reasoning, coding, agents, and visual understanding benchmarks.

Architecture and Key Strengths

The model employs an efficient hybrid architecture combining Gated Delta Networks with sparse Mixture-of-Experts, enabling high-throughput inference with minimal latency and cost overhead. It supports a native context length of 262,144 tokens. Reinforcement learning was scaled across million-agent environments with progressively complex task distributions, enhancing real-world adaptability. The model also expands global linguistic coverage to 201 languages and dialects.

Benchmark Results

In non-thinking mode, Qwen3.5-0.8B achieves 29.7 on MMLU-Pro, 48.5 on MMLU-Redux, 46.4 on C-Eval, 16.9 on SuperGPQA, 52.1 on IFEval, and 34.1 on MMMLU. In thinking mode, results improve to 42.3 on MMLU-Pro, 59.5 on MMLU-Redux, 50.5 on C-Eval, 21.3 on SuperGPQA, 11.9 on GPQA, 44.0 on IFEval, and 21.0 on IFBench. These figures are compared against larger models in the Qwen family, demonstrating competitive performance at a smaller parameter scale.

Model Details

• Parameters: 0.8B
• Hidden dimension: 1024
• Layers: 24
• Context length: 262,144 tokens natively
• Architecture: Gated DeltaNet and Gated Attention with Feed-Forward Networks