Alibaba Group

At the Yunqi Conference, Alibaba laid out a roadmap for Qwen4 and Qwen5, with Qwen4 due soon and later Qwen4.5 and Qwen5 planned to scale toward 5T–10T parameters. The company also described Qwen3.8-Max autonomously completing a chip network-on-chip design with area down 42% and power down 59.5%, and SGLang deployed on a new T-Head Semiconductor GPU in 56 hours, improving throughput by 96%.

Multimodal releases included Qwen-Image-3.1, Qwen-Audio-3.1, the music model HappyShrimp 1.1, and world model HappyOyster 2.0 Preview. Alibaba said a next-generation video model will debut in November and move toward 'Agent Video'; it is also working with universities and industry partners on a world-model benchmark and arena.

Moonshot AI (月之暗面) has closed a pre-IPO private round that values it at about $50 billion (RMB 350 billion), up from $31.5 billion in the summer. It has filed confidentially for a Hong Kong listing planned for the first quarter of next year, aiming to raise up to $5 billion; Bank of America is the overall coordinator, with CICC, Deutsche Bank and Goldman Sachs as sponsor banks. Its backers include Alibaba, Tencent and 5Y Capital, and people familiar say annual recurring revenue is $1 billion, expected to reach $2 billion by December.

Alibaba Group's Qwen-Audio-3.1-TTS has topped Artificial Analysis's Controlled Voice Arena voice leaderboard with an Elo score of 1177 as of September 28, 2026. The model supports multilingual and dialect synthesis, same-timbre cross-lingual transfer, and instruction-based control of emotion, speech rate, and expression. It is part of the Qwen-Audio-3.1 series released at the 2026 Apsara Conference, alongside ASR and Realtime models; APIs for all three are available on the Qwen AI platform, and voice model prices were cut by up to 95%.

At the 2026 Yunqi Conference, Banma Technologies unveiled AutoOmni 2.0-23B-A3B, an on-device omnimodal large model aimed at intelligent cockpits. The sparse mixture-of-experts architecture has 23 billion total parameters and 3 billion active parameters, so the model can run on automotive-grade chips rather than in the cloud. The company says it supports more than eight concurrent tasks, speeds inference by five to six times, keeps quantization fidelity above 99%, and cuts runtime memory use by half. Availability was not specified.

Banma says its earlier 4B dense model improved average cockpit capability by 20%, and the company disclosed coverage of 68% of mainstream automakers and a 63% share in on-device intelligence and end-model applications.