Qwen 3.7 Preview Quietly Lands in Qwen Chat
Yesterday a Qwen 3.7 option showed up in the Qwen Chat dropdown. The Qwen team tweeted a brief preview note. No model card, no weights, no benchmarks. HN sent the thread to 180 points before lunch. This is how Alibaba launches now.
What we know. The chat-only preview is live. Anyone with a Qwen account can poke it. Community testing in the first 24 hours points at sharper reasoning chains, longer instruction-following, and noticeably better tool-use behavior versus Qwen3.6-Max-Preview. No confirmed parameter count, no confirmed context length, no commitment on open weights. Alibaba has been shipping Qwen 3.6 sub-models on a rolling basis since April and the cadence suggests open-weight variants will follow if the chat preview holds up.
The macro frame matters more than the model. Qwen 3.5 dropped in February. Qwen 3.6-Max-Preview in April. Qwen 3.7-Preview in May. Three preview-grade major bumps in under four months while US export controls keep tightening on the chips that train them. This is the rhythm of a team that is sprinting harder, not slower, under sanction pressure. Compare to GPT-5.5 (April) and Claude Opus 4.7 (April). Frontier-class cadence has compressed across all three labs to roughly monthly preview drops with quarterly open-weight or full releases.
The practical question is whether Qwen 3.7 ships open weights. Qwen has been the most open of the top three labs. If a 27B or 35B-A3B variant lands in the next four weeks at the quality the chat preview suggests, the open-source agent stack just got another tier of model with no asterisk. Watch the Qwen GitHub org and HuggingFace this week.
https://chat.qwen.ai
← Back to all articles
What we know. The chat-only preview is live. Anyone with a Qwen account can poke it. Community testing in the first 24 hours points at sharper reasoning chains, longer instruction-following, and noticeably better tool-use behavior versus Qwen3.6-Max-Preview. No confirmed parameter count, no confirmed context length, no commitment on open weights. Alibaba has been shipping Qwen 3.6 sub-models on a rolling basis since April and the cadence suggests open-weight variants will follow if the chat preview holds up.
The macro frame matters more than the model. Qwen 3.5 dropped in February. Qwen 3.6-Max-Preview in April. Qwen 3.7-Preview in May. Three preview-grade major bumps in under four months while US export controls keep tightening on the chips that train them. This is the rhythm of a team that is sprinting harder, not slower, under sanction pressure. Compare to GPT-5.5 (April) and Claude Opus 4.7 (April). Frontier-class cadence has compressed across all three labs to roughly monthly preview drops with quarterly open-weight or full releases.
The practical question is whether Qwen 3.7 ships open weights. Qwen has been the most open of the top three labs. If a 27B or 35B-A3B variant lands in the next four weeks at the quality the chat preview suggests, the open-source agent stack just got another tier of model with no asterisk. Watch the Qwen GitHub org and HuggingFace this week.
https://chat.qwen.ai
Comments