Qwen3.8-Max is a flagship large language model from Alibaba's Qwen team, unveiled as Qwen3.8-Max-Preview on July 19, 2026 at the World Artificial Intelligence Conference in Shanghai and made generally available on August 3, 2026. Alibaba describes the model — 2.4 trillion total parameters, 95 billion activated per query — as "second only to Fable 5," and said at general availability that the weights would follow the next week, which would make it the first model at the company's Max tier to be released openly (Source: bloomberg.com; the-decoder.com; scmp.com).
Two people familiar with the plans said on August 7, 2026 that Alibaba intends to ask major users of the model for a share of the revenue they earn from it, a measure described as due to be implemented the following week, with the rate unsettled and discussions ongoing. The approach follows Moonshot AI's Kimi K3 licence, which requires a commercial agreement from anyone offering the model as a service above $20 million in annual sales (Source: reuters.com).
Announcement and availability
The preview was announced two days after Moonshot AI's Kimi K3 release, which had intensified attention on Chinese open-weight frontier models. The preview is live on Alibaba's Token Plan, Qoder, and QoderWork platforms at 10% of standard price. No license, model card, or benchmark results had been published as of the announcement (Source: the-decoder.com). On July 21, 2026 the Qwen team extended availability to the Qwen PC and web clients alongside Token Plan, Qoder, and QoderWork, positioning the model for full-stack development, data analysis, Office automation, multi-agent systems, and long-horizon tasks, and describing the full Qwen3.8 release as imminent (Source: aidisruption.ai). Alibaba's shares rose as much as 5.4% on July 20, 2026 (Source: bloomberg.com; Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026)).
The promised open-weight release would mark a return to open release at Alibaba's flagship tier after the proprietary, API-only Qwen3.7-Max (May 2026). The announcement followed Xi Jinping's July 18, 2026 speech urging China to "encourage open source, openness, collaboration and sharing" in AI (Source: english.scio.gov.cn; Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026)). See Chinese AI Policy.
Alibaba made the model generally available on August 3, 2026 through Alibaba Cloud's Model Studio APIs and through QwenWork, its workplace agent platform, which entered public beta the same day. The company said the weights would go to Hugging Face and ModelScope the following week, which would make Qwen3.8-Max the first Qwen-Max-class model released openly. Alibaba's Hong Kong-listed shares rose 7% on the day, closing at HK$125.20 (Source: scmp.com). Alibaba priced the model below Moonshot AI's Kimi K3 and said it would publish weights for both Qwen3.8-Max and Qwen3.8-27B the following week. Techmeme's summary of the report puts the pricing at $2 per 1 million input tokens and $6 per 1 million output tokens against Kimi K3's $3 and $15 (Source: theinformation.com); the $2/$6 figures were independently confirmed by The Straits Times, which also reports the intraday share move as up to 7.3% and records Alibaba's own positioning — comparable or in places better scores than Fable 5, and higher than Kimi K3 on several benchmarks — alongside the claim that the model "activates only some parts when in use to reduce computational cost and latency" (Source: straitstimes.com).
Reported architecture and capabilities
Qwen developer Shuai Bai described Qwen3.8 as the team's first multimodal model above one trillion parameters, processing text, images, video, and documents, and said it should improve on Qwen3.7-Max in coding, full-stack development, data analysis, and office workflows. Alibaba's own announcement called the model "one of the most powerful available today, comparable to leading frontier systems." No benchmark table, model card, license, or active-parameter count accompanied the preview, so the performance claims stand as the developer's positioning rather than measured results; no independent evaluator had scored the model at the time of the preview (Source: marktechpost.com; scmp.com).
The undisclosed active-parameter count was the principal open question about the model's practical cost at preview. Alibaba's Max tier uses sparse mixture-of-experts designs — Qwen3-235B-A22B carries 235 billion total parameters while activating 22 billion per token, and Qwen3-30B-A3B activates roughly 3 billion — so the headline 2.4-trillion figure constrains serving cost only loosely. Commentary at the time of the preview noted that a 2.4-trillion-parameter model at 4-bit precision would require roughly 1.2 terabytes for weights alone, against 141 GB of memory on a single Nvidia H200, which makes the availability of a smaller activated-parameter variant, a quantised checkpoint, or a distilled sibling the more consequential question for anyone intending to run the weights locally (Source: marktechpost.com). The figure was disclosed at general availability: the model activates 95 billion parameters per query of its 2.4 trillion total, and supports a 1 million-token context window (Source: the-decoder.com).
Alibaba published two case studies with the general-availability release, both directed at long-horizon autonomous operation. In the first, the model spent 16 days building the command-line tool oh-my-cli, recording 265 commits, 127 pull requests and 151 issues by July 30, 2026 without human intervention. In the second, it reduced a cryptographic circuit from 8,298 logic gates to 678 over roughly 500 iterations. Alibaba's self-reported benchmarks put the model at 93 on PaperBench and 86.6 on TerminalBench 2.1, the latter behind GPT-5.6 Sol's 88.8. Independent verification of these figures was pending at release (Source: the-decoder.com).
Reaction among developers was described as cautiously positive and split along two lines: support for continued open-weight competition among Chinese labs, set against fatigue over unverified benchmark claims and doubts about who could serve a 2.4-trillion-parameter model in practice. Some commentators read the timing as a direct response to Kimi K3, a 2.8-trillion-parameter open-weight model released two days earlier, and questioned the "second only to Fable 5" framing (Source: marktechpost.com).
Relationships
- instance-of: Open-Weight Frontier Models (pending the promised weights release), General-Purpose AI (GPAI), Mixture of Experts (MoE)
- developer: Alibaba / Qwen Team
- related: Qwen3, Kimi K3 — the release it was widely read against; GLM-5.2, DeepSeek V4 Pro / V4 Flash — peer Chinese open-weight flagships; Chinese AI Policy, US-China AI Competition: Different Races, Different Metrics, Inference Economics and Token Pricing