Qwen/Qwen2-1.5B-Instruct

Qwen2-1.5B-Instruct is a small, text-only instruction model with unusually accessible deployment support: the official card documents Transformers, vLLM/OpenAI-compatible serving, SGLang, and Docker, under Apache-2.0. Its 1.5B scale supports low-cost self-hosting and broad local availability, but materially limits general capability. It is well below Gemini 3.5 Flash and GPT-5.6 Luna on technical breadth and multimodal I/O, and far below Claude Opus 4.8 or GPT-5.5 for advanced reasoning and agentic work. Coding evidence is weak rather than absent: the Qwen2.5 report lists Qwen2-1.5B at 4.5 on LiveCodeBench, while the displayed v5 leaderboard has no exact-model entry. DeepSWE has no row, no SWE-bench direct result was found, and the supplied Terminal-Bench/Aider sources establish ecosystem presence rather than benchmark performance. The ACL evaluation’s 1.8% Arena-Hard win rate against GPT-4 (2.8% style-controlled) is a strongly negative preference signal. These results support scores below Pegasus 1.5’s 35 coding score despite Qwen’s stronger general developer tooling. Apache-2.0 weights and standard serving paths make it much more cost-effective and accessible than proprietary frontier anchors, but no model-specific hosted pricing was supplied, reducing pricing clarity. Qwen has meaningful provider-level public adoption and distribution, yet exact-model adoption is not quantified. Evidence quality is moderate: official model documentation and a technical-report comparison are useful, but independent benchmark coverage is sparse and Artificial Analysis evidence supplies no stated metrics here.