Alibaba

Alibaba is the Chinese hyperscaler behind the Qwen family of models, one of the most prolific release cadences in AI. Their May 2026 flagship Qwen3.7-Max landed on the Alibaba Cloud API on May 19 and was formally unveiled at the 2026 Alibaba Cloud Summit on May 20. It carries a 1 million token context window, an extended-thinking mode, and posted the top result on the public Artificial Analysis Intelligence Index at 57, with roughly 1,475 Elo on the LM Arena text leaderboard. Alibaba claims Qwen3.7-Max can run agentic workloads autonomously for up to 35 hours on long-horizon tasks. On August 3, 2026 Alibaba replaced it at the top with Qwen3.8-Max, a 2.4 trillion parameter MoE activating roughly 95 billion per token, priced at $2 input / $6 output per 1M with a 1M context window and native text, image, and video input. Its benchmark profile is split: 86.6 on Terminal-Bench 2.1 (ahead of Claude Opus 4.8 and Fable 5 at 84.6, behind GPT-5.6 Sol at 88.8), leading PaperBench at 93.0 and IFBench at 82.8, but well behind Fable 5 on core software engineering at 67.7 versus 80.0 on SWE-bench Pro. Eleven days later, on August 14, Alibaba published Qwen3.8-27B under Apache 2.0: 27.78 billion dense parameters, natively multimodal, 262K context, and the strongest locally deployable multimodal model near that size. On August 26 the team published Qwen3.8-Flash-Next, an open-weight preview of the Qwen4 architecture: 125 billion parameters with only 6 billion activated per token, plus a 51 billion parameter n-gram embedding table that can live in system RAM rather than GPU memory, and 262K native context extensible to 1M with YaRN. Alibaba claims it beats the 397 billion parameter Qwen3.7-Plus at roughly one ninth the training cost. The managed build ships as Qwen3.8-Flash on QwenCloud at $0.15 input and $0.47 output per 1M tokens (it launched at $0.16) with a 1M default window, which is about a twelfth of the Qwen3.8-Max rate on both sides. Earlier members of the family including Qwen3 Coder Next, Qwen3.5 Plus, and Qwen3.6 Plus remain widely deployed across OpenRouter and self-hosted setups. On September 5, 2026 the qwen3.8-max endpoint moved to qwen3.8-max-0902, a September 2 post-training snapshot, with pricing unchanged. On September 18 the team added Qwen3.8-Omni-Flash, its first omni-modal model built around agentic work: text, image, audio, and video in, text only out, on the same $0.15 and $0.47 rate card as Qwen3.8-Flash with a 1M context window and implicit cache reads at $0.016. The per-token price is not the story there. Alibaba says an hour of audio input now costs 98 percent less than on Qwen3.5-Omni-Plus and an hour of audio and video more than 93 percent less, and reports an average gain above 25 percent across 29 evaluations. The catch is that the Omni name no longer implies speech output: this model returns text only, so voice applications still need a separate synthesis stage.