Alibaba officially announced Qwen 4 at its Apsara Conference. The announcement is newsworthy to practitioners tracking the Qwen model family's next generation, though the posts contain no further technical specifics.
A Reddit user says a self-described 0-day partner of Alibaba casually indicated that Qwen 4 is planned for release at the end of October, sooner than the user expected given the gap between Qwen 3.6 and 3.8. The claim is unverified rumor, acknowledged by the poster as 'in bro we trust.' Practitioners may treat it as a soft signal, not confirmation.
A Reddit user says a self-described 0-day partner of Alibaba casually indicated that Qwen 4 is planned for release at the end of October, sooner than the user expected given the gap between Qwen 3.6 and 3.8. The claim is unverified rumor, acknowledged by the poster as 'in bro we trust.' Practitioners may treat it as a soft signal, not confirmation.
Alibaba announced the Qwen 4 family at its Apsara conference, with a Medium writeup separating confirmed details from rumor. The open-weights tier is described as coming "very soon," and the architecture reportedly already runs on consumer gaming GPUs.
Alibaba announced the Qwen 4 family at its Apsara conference, with a Medium writeup separating confirmed details from rumor. The open-weights tier is described as coming "very soon," and the architecture reportedly already runs on consumer gaming GPUs.
Qwen released Qwen3.8-Flash-Next, an open-weight hybrid model that previews the Qwen 4 architecture. The 125B MoE model reportedly outperforms DeepSeek V4 Flash at half the parameters and is stronger than Opus 4.6. Its use of n-gram tables could enable running 1T+ parameter models on a single server with modest GPUs and system RAM.
Qwen released Qwen3.8-Flash-Next, an open-weight hybrid model that previews the Qwen 4 architecture. The 125B MoE model reportedly outperforms DeepSeek V4 Flash at half the parameters and is stronger than Opus 4.6. Its use of n-gram tables could enable running 1T+ parameter models on a single server with modest GPUs and system RAM.
Alibaba released Qwen3.8-Max, its largest and most capable AI model to date, a 2.4T-parameter mixture-of-experts model focused on coding, agents, and professional cowork tasks. The model is available via API at $2/M input and $6/M output tokens, with open weights expected next week. Alibaba claims performance rivaling top US frontier labs and domestic rivals like Moonshot AI's Kimi K3.
Alibaba released Qwen3.8-Max, its largest and most capable AI model to date, a 2.4T-parameter mixture-of-experts model focused on coding, agents, and professional cowork tasks. The model is available via API at $2/M input and $6/M output tokens, with open weights expected next week. Alibaba claims performance rivaling top US frontier labs and domestic rivals like Moonshot AI's Kimi K3.
Alibaba's Qwen team posted a cryptic teaser on X (formerly Twitter) on July 19, 2026, hinting at an upcoming model release. The post, shared on Reddit, has garnered community attention and speculation about the next iteration of the Qwen series.
Alibaba's Qwen team posted a cryptic teaser on X (formerly Twitter) on July 19, 2026, hinting at an upcoming model release. The post, shared on Reddit, has garnered community attention and speculation about the next iteration of the Qwen series.