qwen3.5-2b

Qwen3.5-2B is the small-scale member of the Qwen3.5 family, distributed as a post-trained model whose weights are published on Hugging Face for prototyping, task-specific fine-tuning, and other research or development purposes. The family is positioned around a unified vision-language foundation that trains on multimodal tokens via early fusion, aiming for cross-generational parity with Qwen3 and improved performance over Qwen3-VL models on reasoning, coding, agents, and visual understanding. The 2B-class variant on Ollama confirms roughly 2.27B parameters in the qwen35 architecture, packaged with Q8_0 quantization at about 2.7GB and released under Apache License 2.0, with capability tags indicating vision input along with reasoning and tools support.

Because of its modest size, Qwen3.5-2B is best suited to experimentation, lightweight assistants, and on-device or cost-sensitive deployments rather than heavy production reasoning. It inherits the broader Qwen3.5 emphasis on efficient hybrid inference, scalable reinforcement learning across million-agent environments, and broad linguistic reach, making it a flexible base for fine-tuning into specialized domains. The combination of multimodal grounding, reasoning support, and tool-oriented behavior offers a practical entry point for developers who want to explore Qwen3.5's design without committing to the larger tiers, while still benefiting from the family's vision-language and global language advances.

Requestyqwen3.5-2bqwen3.5

Open provider page Provider docs

Quick Info

Provider: Requesty
Model key: qwen3.5-2b
Release date: Mar 10, 2026
Last updated: Mar 10, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.02
Output token cost: $0.10

Limits

Output tokens: 262,144 tokens
Context window: 262,144 tokens

Latest news about qwen3.5-2b

Requesty

Coverage

Qwen3.5-2B is a 2-billion parameter post-trained causal language model with a vision encoder, released by the Qwen team in the Qwen3.5 family. The Hugging Face repository distributes weights and configuration files compatible with Transformers, vLLM, SGLang, and KTransformers, intended for prototyping, task-specific fine-tuning, and research or development use at small scale. Architecture details include 24 layers, a 2048 hidden dimension, and a padded token embedding of 248,320. The model introduces a hybrid layout combining Gated DeltaNet linear attention with sparse Mixture-of-Experts and gated attention, using 6 blocks of 3 (Gated DeltaNet -> FFN) followed by 1 (Gated Attention -> FFN). Qwen3.5 highlights cited on the card include a unified vision-language foundation, scalable reinforcement learning across million-agent environments, and expanded coverage of 201 languages and dialects. The card states Qwen3.5 targets exceptional utility and performance with multimodal learning, architectural efficiency, and next-generation training infrastructure.

Requesty

Coverage

Ollama distributes qwen3.5:2b as a 2.27B-parameter Q8_0 quantized package (~2.7GB) under Apache License 2.0, mirroring the upstream Qwen3.5-2B card content for local multimodal inference. The page shows 20.9M downloads and indicates the model supports vision, tools, thinking, and cloud modes for multimodal workflows. Repository metadata lists the qwen35 architecture and an updated tag, confirming independent downstream packaging of the 2B variant. The Ollama listing reproduces the Qwen3.5 family highlights, including Gated Delta Networks paired with sparse MoE for efficient inference, multimodal training efficiency approaching text-only levels, and support for 201 languages. A benchmark table on the page compares Qwen3.5-397B-A17B and sibling variants against GPT5.2, Claude 4.5 Opus, Gemini-3 Pro, Qwen3-Max-Thinking, and K2.5-1T-A32B across MMLU-Pro, MMLU-Redux, SuperGPQA, C-Eval, IFEval, and IFBench. These sibling-model comparisons illustrate the family's positioning rather than 2B-specific scores.

Requesty

Official sourceBenchmark

Requesty's master model catalog positions Qwen3.5-2B within a broader lineup of 210 unique models across 33 providers and 700 endpoints, all reachable through a single OpenAI-compatible API. The page confirms Qwen3.5 family variants from Alibaba (Qwen) are available on the platform, alongside models from OpenAI, Anthro The catalog provides at-a-glance pricing and context for multiple Qwen3.5 variants (including flash, max, and MoE configurations like `qwen3.8-2.4t-a95b` and `qwen3.8-max`), helping developers compare Qwen3.5-2B against other Alibaba offerings on Requesty. It serves as a discovery surface that links to per-model pages