Alibaba's Qwen releases Qwen3.7-Plus multimodal agent model
Alibaba's Qwen team has released Qwen3.7-Plus, which combines vision and language in one agent model and leads its comparison table on Terminal Bench 2.0, while trailing Claude Opus 4.6 on SWE-bench Verified.
Alibaba's Qwen team introduced Qwen3.7-Plus, a multimodal model that adds upgraded vision-language capabilities to the Qwen3.7 text backbone. Qwen calls it a "multimodal interactive hybrid agent." According to the team, it can read screens and operate GUIs, write code from visual references, navigate mobile apps from start to finish, and combine GUI and CLI actions in a single agent loop. Qwen also says it performs consistently across agent scaffolds, including Claude Code, OpenClaw and Qwen Code.
In Qwen's text benchmarks, the model has the highest score in its comparison table on Terminal Bench 2.0. It scores 70.3, ahead of DeepSeek-V4-Pro Max at 67.9 and Opus-4.6 Max at 65.4. It also leads the table on Deep-Planning (62.3 vs 58.9 for Opus), MCP-Mark (58.7) and MRCR-v2 at 128k context (91.7 vs 84.0).
The results are mixed elsewhere. On SWE-bench Verified, Qwen3.7-Plus scores 77.7, below Opus at 80.8 and below its predecessor, Qwen3.6-Plus, at 78.8. It also trails Opus on HLE (34.7 vs 40.0), ClawEval and CoWorkBench. Several of the benchmarks are Qwen's own, including QwenWebDev, QwenSVG, QwenWorldBench and the internal CoWorkBench. Qwen ran the SWE-bench tests on an internal scaffold and corrected some SWE-Pro tasks before re-scoring all baselines.
The material provided doesn't include vision benchmark results, pricing, availability or whether the weights will be released.