Qwen 3.7 Debuts: Ranks 13th Globally and Tops China’s Model Leaderboard

On May 18, the Arena official account announced the latest scores for Qwen 3.7‑Max‑Preview and Qwen 3.7‑Plus‑Preview. The Qwen official account replied that the Qwen 3.7 series is about to be released, noting that the previous version was still being discussed when the next one arrived.

According to Arena’s leaderboard, Qwen 3.7‑Max‑Preview ranks 13th overall in the text domain, making it the only Chinese model among the top‑15 and lifting Alibaba’s lab to 6th place globally. The models ahead of it are Claude Opus 4.6/4.7, Gemini‑3.1/3 Pro, and GPT‑5.4 series. Qwen 3.7‑Plus‑Preview ranks 16th in the vision domain, giving Alibaba’s lab a top‑5 position on the visual track and the highest rank for a domestic model on that list.

In specialized tracks, Qwen 3.7‑Max‑Preview shows strong depth: 7th in Mathematics, 9th in Expert‑Prompt, 9th in Software/IT, and 10th in Coding, all within the global top‑10. The author emphasizes that the model improves not only overall capability but also makes breakthrough progress in mathematical reasoning, domain knowledge, and code generation.

Accelerated Model Release Cadence

Qwen 3.6‑Max‑Preview was released at the end of April; by May 19 the Qwen 3.7 preview was already on the table. This “two‑generation parallel iteration” is rare in the industry.

Historically, Qwen’s major releases in 2023‑2024 followed a 4‑to‑6‑month cycle, similar to mainstream industry pacing. The turning point came with the Qwen 3 series (2025), which introduced Dense and MoE models ranging from 0.6 B to 235 B parameters and a dual‑mode design: a “Thinking” mode for complex reasoning, long‑chain decisions, and agent tasks, and a “Non‑Thinking” mode for low‑latency responses.

From 2026 onward, the cadence sped up further, with major versions arriving every 2‑3 months and almost monthly updates. The typical workflow is preview release → community testing → developer adoption → final release, enabling rapid market feedback and a “small‑step, fast‑run” iteration strategy.

Domestic Competition Heats Up

Since 2025, China’s large‑model race has shifted from pure technical competition to ecosystem competition among Alibaba, ByteDance, Tencent, and Baidu. Capital expenditure of the four giants grew 45% in 2025 and is projected to rise another 30% in 2026, indicating a battle for ecosystem control rather than raw model ability.

Other notable players include DeepSeek (which will launch DeepSeek‑V4 in 2026), Tencent’s newly restructured HunYuan Hy3 Preview, and Xiaomi’s MiMo v2.5 Pro (ranked 7th on OpenRouter). Zhipu released GLM‑5 and GLM‑5.1, with the latter achieving SOTA performance in coding‑agent scenarios and raising API prices to near‑Anthropic levels; despite an 83% price hike, API calls grew 400% in Q1 2026.

MiniMax introduced the flagship M2.7 model in March, showcasing a “self‑evolution” path via the Agent Harness framework, delivering 30‑50% workload reduction and ~30% performance gain on internal benchmarks. It also launched Music 2.6, cutting first‑packet latency to under 20 seconds.

Kimi’s backing organization, Moonlight, released the open‑source Kimi K2.6 (1 T‑parameter MoE) in April, supporting 13‑hour continuous coding and 5‑day autonomous agent runs, reclaiming the global open‑source model lead. Its annual recurring revenue surpassed $200 million in April 2026, accompanied by a $2 billion financing round and a post‑money valuation exceeding $200 billion.

Other firms such as ZhiPu, MiniMax, Kimi, ZhiShu, SenseTime, Ant Group, and others continue to launch models, intensifying the competition.

Overall, by 2026 the Chinese large‑model arena’s focus has moved to commercial efficiency and ecosystem control, with success now hinging on who can close the commercial loop fastest within the “open‑source ecosystem + pricing strategy + customer stickiness” triangle.

Reference: https://x.com/Alibaba_Qwen/status/2056403591464984753