Alibaba / Qwen Team

The Qwen team is Alibaba Cloud's foundation-model group. It has shipped a multi-generation family of open-weight large language models and is, together with DeepSeek, one of the two most prolific Chinese producers of frontier-competitive open-weight models. Its May 2025 Qwen3 release applied an Apache 2.0 license to a 235B-parameter mixture-of-experts (MoE) flagship, one of the most permissive licenses applied to a frontier-scale model to that point (Source: Qwen3 Technical Report). In May 2026 Alibaba departed from that posture for its top tier, releasing the agentic flagship Qwen3.7-Max as a proprietary, API-only model.

Snapshot

Financials and AI capital spending

Models

Qwen3.7-Max claimed performance (Alibaba figures)

The ~35-hour, 1,158-tool-call autonomous-run figures are Alibaba's own claims and are not yet independently verified; confidence on these figures is medium.

Overview and relationship to Alibaba Cloud

The Qwen team sits inside Alibaba Cloud and serves both as an AI capability for Alibaba's own products and as a model-as-a-service offering. Its open-weight releases complement, rather than replace, Alibaba Cloud's hosted API. This posture differs from OpenAI and Anthropic (closed weights, API-only) and from Meta (open weights, but not a public-cloud customer-serving posture).

Financing and AI capital expenditure

Alibaba's AI spending moved from a stated multi-year plan to a visible drag on reported earnings during 2026. Reporting April–June results on August 20, 2026, the company said quarterly net profit fell 75% year over year as AI capital expenditure rose 75% to ¥67.68 billion. Revenue rose 9% to ¥268.95 billion against an LSEG consensus of ¥268.88 billion; AI cloud and compute services revenue rose 45% to ¥48.44 billion, and the model-as-a-service business passed ¥16 billion in annual recurring revenue. Adjusted earnings of ¥8.52 per American depositary share missed an estimate of ¥10.53, and the US-listed shares fell 4.6% in early trading. Chief executive Eddie Wu said the company had already spent half of its planned ¥380 billion (about $56.4 billion) 2026–2029 AI investment within the year, that it expects to break even on AI-related capital expenditure within three years at current average gross margins, and that chips from its in-house T-Head unit are being deployed at scale on "supernodes" — server racks linking hundreds of semiconductors — for both training and inference (Source: reuters.com).

Three days later, on August 23, 2026, the company launched a HK$80 billion (US$10.2 billion) Hong Kong share placement to fund AI development — reported as the largest primary follow-on offering by a Hong Kong-listed company and the world's third-largest primary follow-on sale of the year, after Alphabet's and Intel's. A term sheet reviewed by Reuters priced 710 million ordinary shares at HK$112.70, a 3.6% discount to the most recent close. Alibaba said it intends to use all net proceeds on its "full stack" AI capabilities, which it defined as chips, infrastructure, and model development and deployment, without giving a breakdown by category. Two people familiar with the deal said sovereign wealth funds were among the buyers and that the offering was increased after being oversubscribed; Morgan Stanley, HSBC, UBS and CICC are joint bookrunners. The placement was not registered under US securities laws, making American investors ineligible to participate (Source: reuters.com). See AI Bubble Debate, Sovereign AI.

Product family

The Qwen line runs from the early open-weight Qwen and Qwen1.5 dense models through Qwen2 (expanded scale and multilingual support) to Qwen2.5, which added 29-language support and spawned specialist variants Qwen2.5-VL (vision-language), Qwen2.5-Math, and Qwen2.5-Coder (Source: Qwen3 Technical Report). The May 2025 Qwen3 release comprised 8 models from 0.6B to 235B parameters, with unified thinking/non-thinking modes, an adjustable thinking budget, and 119-language coverage, all under Apache 2.0 (Source: Qwen3 Technical Report). In May 2026 Alibaba added Qwen3.7-Max, a long-horizon agentic flagship released proprietary and API-only.

Specialist Qwen2.5 variants were used as synthetic-data generators for Qwen3 pre-training: Qwen2.5-VL for PDF text extraction, and Qwen2.5-Math and Qwen2.5-Coder for domain-specific synthetic data (Source: Qwen3 Technical Report). This is a self-improvement loop inside the Qwen family, comparable to the R1→V3 distillation inside DeepSeek (see DeepSeek-V3 Technical Report).

Qwen3.7-Max

Alibaba released Qwen3.7-Max on May 21, 2026 as an agentic, long-horizon flagship. Alibaba says the model ran for roughly 35 hours of continuous autonomous execution, comprising 1,158 tool calls and 432 kernel evaluations, optimizing code on hardware unseen during training, and scored 44.5 on the Apex Math Reasoning benchmark against Claude Opus 4.6 Max's 34.5 (Source: venturebeat.com). The model is designed to support external agent harnesses, including Anthropic's Claude Code (Source: venturebeat.com).

Qwen3.7-Max was released proprietary and API-only, departing from Alibaba's prior open-weight releases such as the Apache 2.0 Qwen3 family. The closed release of the flagship tier indicates Alibaba is treating its most capable agentic model as a commercial API asset rather than an open-weight release, while the open Qwen3 family persists at lower tiers. This bears on the question, discussed in Open-Source AI / Open-Weight Models and China and the US Are Running Different AI Races, of whether Chinese labs are tilting toward closed releases (see Licensing and policy posture, below).

Qwen3.8-Max preview

On July 19, 2026, the Qwen team unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter flagship it describes as "second only to Fable 5," at the World Artificial Intelligence Conference in Shanghai, two days after Moonshot's Kimi K3 release (Source: bloomberg.com; the-decoder.com). Open weights are promised "soon" — a return to open release at the flagship tier after the closed Qwen3.7-Max — and the preview is live on Alibaba's Token Plan, Qoder, and QoderWork at 10% of standard price, though no license, model card, or benchmarks had been published at the preview announcement (Source: the-decoder.com). Alibaba's shares rose as much as 5.4% on July 20 (Source: bloomberg.com; Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026)). The announcement followed Xi Jinping's July 18 speech urging China to "encourage open source, openness, collaboration and sharing" in AI, which Stratechery's Ben Thompson read as preceding Alibaba's return to open-weight releases (Source: english.scio.gov.cn; Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026)). See Chinese AI Policy.

Qwen 3.8 27B

The Qwen team released Qwen 3.8 27B on Friday, August 14, 2026: an Apache 2.0-licensed, 27-billion-parameter vision-capable model that ships defaulting to the xhigh reasoning-effort setting. Testing the 17GB Q4_K_M build on an M5 Max MacBook Pro and an NVIDIA DGX Spark, Simon Willison reported on August 16 that a single SVG drawing took 21 minutes and 22,276 reasoning tokens to produce 3,223 output tokens at that default, against 3,715 tokens in 137 seconds with reasoning turned off, and measured 15 to 30 tokens per second in LM Studio (Source: simonwillison.net). The release continues the open-weight line at the mid tier alongside the flagship Qwen3.8-Max.

Alibaba's statement that the model had been downloaded more than 1 million times within a few days of release became public on August 18, 2026, alongside its description of the model as one of the company's fastest-growing and as an open-source system that accepts text, image and video prompts and runs locally on a personal computer. Only the briefing's abstract was retrievable, so the download figure's measurement basis and the platforms it covers are not established here (Source: theinformation.com).

Ecosystem scale and download figures

Alibaba said in an emailed statement reported on August 15, 2026 that Qwen has open-sourced more than 460 models, that its ecosystem has spawned more than 300,000 derivatives, and that its open-weight models drew more than 3 billion global downloads in the past six months, against 418 million for Google and 227 million for Meta in 2026 on Hugging Face's figures (Source: bloomberg.com).

Hugging Face's own State of Open Models report of August 14, 2026 counted 151,448 Qwen-based derivatives on its Hub, 2.6 times Meta's total footprint and ahead of Google at 82,506 (Source: huggingface.co). The two derivative counts differ by roughly a factor of two and neither source reconciles them: Hugging Face's is Hub-only and independently measured, while Alibaba's is ecosystem-wide and self-reported.

Nathan Lambert attributes China's overtaking of the United States in Hugging Face downloads of open models, which he dates to July 2025, "primarily" to the success of Qwen. On his own tracking, China's cumulative download lead had grown to about 1.6 billion by September 2026, on a Chinese total of 3.2 billion, twice the American total. This is a country-level cumulative Hugging Face figure; it is not directly comparable with Alibaba's self-reported count of more than 3 billion global downloads for Qwen alone over six months, and neither source addresses the other. Lambert also writes that "to a first order approximation, most of academic research is conducted on" Qwen: in his scan of five machine-learning categories of arXiv, Qwen was mentioned in about 7.5% of papers in the fall of 2024 and 30% in September 2026, against about 21% for Meta's Llama. He reports that members of Qwen's leadership he met in China are "very invested in and intentional about this type of adoption" (The current balance of power in open models (Nathan Lambert, September 2026)).

Scale-up roadmap and T-Head chips

At Alibaba Cloud's Apsara Conference in Hangzhou on September 22, 2026, chief executive Eddie Wu said Alibaba plans to train a model of 5 trillion to 10 trillion parameters, two to four times the 2.4 trillion of its flagship Qwen 3.8 Max, with Qwen 4 already in training and the scale-up projected for the Qwen 4.5 and Qwen 5 series. He said the Qwen team had made "meaningful progress" on recursive self-improvement, and set a target of more than 20 gigawatts of global data-centre capacity by 2032 (Source: reuters.com). The conference introduced the Zhenwu V900 chip from Alibaba's T-Head unit, which Alibaba says delivers three times the performance of the Zhenwu M890 released in May, with 216 GB of memory, 1,200 GB/s of chip-to-chip bandwidth and native FP8 and FP4 support, and is scheduled for mass production in the first quarter of 2027 (Source: techwireasia.com). Further detail on the keynote is on Wu's page. Alibaba Cloud also said at the conference that it would open its first cloud regions in Turkey, Finland and the Netherlands over the following 12 months, with the Netherlands region first, in October (Source: bloomberg.com; marketscreener.com; aiweekly.co).

Alibaba's Hong Kong-listed shares fell 4.4% on September 23, 2026, alongside falls for Z.ai and MiniMax, after reports that China's Cyberspace Administration was investigating potential data leaks to Anthropic (Source: theinformation.com).

Chip software stack

Alibaba open-sourced its chip software stack in a move reported as targeting Nvidia's dominant CUDA ecosystem, per reporting circulating July 19, 2026 (Source: scmp.com). See Semiconductor Supply Chain.

Licensing and policy posture

Qwen3 is licensed under Apache 2.0, which permits commercial use, modification, and redistribution (Source: Qwen3 Technical Report). Its 119-language coverage and permissive license make it relevant to AI Sovereignty debates, since non-aligned states and multilingual deployments can use Qwen3 without US API dependency.

Alibaba's frontier-scale Apache 2.0 release runs against the framing in Open-Source AI / Open-Weight Models and the Epoch AI account (Source: epochai.substack.com) that "Alibaba has recently tilted back toward closed releases"; as of May 2025 Alibaba was still shipping its frontier-scale flagship open. The May 2026 proprietary, API-only release of Qwen3.7-Max moves in the opposite direction at the flagship tier, while the open Qwen3 family continues at lower tiers, leaving Alibaba's posture bifurcated between an open mid-tier and a closed flagship.

In the broader US-China competition, Alibaba and DeepSeek together anchor the pattern often summarized as "China ships open, US ships closed" at the frontier (China and the US Are Running Different AI Races), with Qwen3's Apache 2.0 release at 235B as a strong data point and the Qwen family serving as a counterexample to claims that Chinese labs are moving uniformly toward closed releases (Open-Source AI / Open-Weight Models). Strong-to-weak distillation inside the Qwen family parallels the efficiency strategies documented at DeepSeek (Distillation, Fast-Follow Problem). The combination of Apache 2.0 licensing and 119-language coverage positions Qwen3 as a substrate for national and regional AI deployments outside the US or China (AI Sovereignty).

A further shift in the commercial terms attached to the open tier became public on August 7, 2026, when two people familiar with the plans said Alibaba intends to ask major users of the next version of its Qwen open-source model for a share of the revenue they make from it. The measure was described as due to be implemented the following week and as applying to Qwen3.8-Max, whose learned settings are available for download. To that point Alibaba had charged developers for use of its models hosted on its own cloud platform while allowing most of its open-source offerings to run in customers' own data centres without payment. The rate had not been settled, with discussions ongoing (Source: reuters.com). The approach follows the licence Moonshot AI applies to Kimi K3, and would narrow the practical difference between an open-weight release and a licensed one for the largest downstream users; see Open-Weight Frontier Models.

On June 24, 2026 Anthropic publicly accused Alibaba of "brazenly" and "illicitly" attempting to extract its AI capabilities, in a June 10 letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs (Senators Tim Scott and Elizabeth Warren) first reported by Bloomberg. Anthropic called it "the largest known distillation attack on Anthropic to date," alleging operators it tied to Alibaba and its AI lab conducted 28.8 million model exchanges through roughly 25,000 fraudulent accounts between April 22 and June 5, 2026. Alibaba had not publicly responded as of the reports; the accusation extends the adversarial-distillation framing previously directed at DeepSeek, Moonshot, and MiniMax (Source: cnbc.com; bloomberg.com).

Alibaba banned employees from using Anthropic's Claude Code and ordered Claude models removed from work computers, according to reporting on July 3, 2026, citing alleged backdoor risks and directing staff to its own Qoder coding platform. The ban, which takes effect July 10, 2026 and classifies Claude Code as high-risk software, followed the June distillation accusation and developer reports that Claude Code inspected user environments — behavior an Anthropic employee described on June 30 as an anti-abuse "experiment we launched in March" (Source: reuters.com; theinformation.com). Anthropic's Thariq Shihipar acknowledged on July 4 that the March experiment could identify Chinese users, saying it targeted reseller abuse and distillation and that stronger mitigations had since been deployed (Source: techcrunch.com).

Alibaba and Tencent backed Kuaishou's Kling AI video unit in a $2.8 billion fundraise announced July 3, 2026 (Source: reuters.com).

Alibaba was among the firms — with ByteDance and Z.ai — that Chinese authorities met over the month before July 7, 2026 about restricting overseas access to China's most advanced AI models, including unreleased ones; the Ministry of Commerce-led discussions also covered criminalizing theft or leaks of proprietary AI technology and new rules on funding domestic AI startups (Source: reuters.com). See Export Controls (AI).

Alibaba is the addressee of the only published copy of the letters in which Rep. Ro Khanna, ranking member of the House Select Committee on the CCP, asked Chinese developers to join a frontier-pacing agreement. Khanna wrote to chief executive Eddie Wu and chairman Joe Tsai in a letter dated September 16, 2026, urging "China's leading artificial intelligence (AI) companies to enter into a binding, enforceable, and verifiable agreement alongside America's leading artificial intelligence (AI) labs to pace the frontier." The letter asks whether Alibaba would agree to slow its rate of AI advancement, execute such an agreement "including by adopting common sense safeguards like global limits on unchecked recursive self-improvement," and accept "neutral third party auditing and all-access inspections," and, failing that, what safeguards it has "to prevent the loss of control or misalignment" of its models; written answers were requested by September 22, 2026. It also states that PRC AI models have facilitated human rights abuses and help modernize the People's Liberation Army (Khanna letter to Alibaba on pacing the frontier (September 16, 2026)). The Washington Post, which published the letter on September 17, reported parallel letters to Moonshot AI and DeepSeek and that Chinese officials and technology leaders had so far rejected calls to slow AI development (Source: washingtonpost.com). No Alibaba response had been reported as of September 18, 2026. See Coordinated slowdown proposals compared.

Enterprise adoption and the data-security debate

On July 15, 2026, Alibaba confirmed that its Qwen model will power Apple Intelligence in China after the Cyberspace Administration of China added Apple's AI services to its approved list; Alibaba's U.S.-listed shares rose 4% in premarket trading on the confirmation (Source: cnbc.com). The integration concludes the Apple-Alibaba China AI rollout first reported in 2025 and delayed amid US-China trade tensions.

In a Bloomberg TV interview on May 20, 2026, Airbnb CEO Brian Chesky defended Airbnb's use of Qwen for its customer-service chatbot and contested the framing that Chinese-developed open-weight models pose data-security risks, stating: "We are not providing data to any Chinese companies. They don't have access to any data… An open-source model does not have access to data. It doesn't work that way." Chesky said Airbnb is not a customer of Alibaba, meaning it uses Qwen open weights rather than Alibaba Cloud's hosted API, and that Airbnb primarily uses a variety of open-source models, including US open-source models. Airbnb's AI customer-service agent handles 70% of customer-service tickets (Source: bloomberg.com).

Chesky's statement was a CEO-level public articulation of the argument that open weights can be data-isolated, made in response to an April 29, 2026 probe by US House committees on China and Homeland Security into Airbnb and Anysphere/Cursor over their use of Chinese AI models. The committee letter cited national-security and data-security concerns about Airbnb's stated preference for "fast and cheap" Qwen in an October 2025 Chesky interview (Source: bloomberg.com).

In a Google I/O fireside chat the same week, Sundar Pichai voiced a similar view of open-weight provenance: "If it is open source with the right licenses, it should matter less where it came from… I worry less about, 'Are we adopting open source models from China?' and more, 'Are we doing enough in the US to make sure we are staying at the frontier?'" (Source: bloomberg.com).

Adoption of Chinese open-weight models extends beyond Airbnb. For May 2026, seven of the ten most-used models on OpenRouter, a multi-million-developer model-routing platform, were developed by Chinese companies, including DeepSeek, Moonshot AI, Tencent, and Qwen (Source: openrouter.ai; cited via bloomberg.com).

People

At the group level, Eddie Wu is chief executive officer and Joe Tsai is chairman of Alibaba Group Holding Limited, as named in the September 2026 congressional letter addressed to them (Khanna letter to Alibaba on pacing the frontier (September 16, 2026)).

The Qwen3 report lists approximately 60 authors. Individual leadership attributions are not yet independently corroborated; specific names will be added when confirmed by additional sources. Contact details are not reproduced.

Sources and related pages

The primary technical source is Qwen3 Technical Report (arXiv 2505.09388, May 2025), the main technical reference for Qwen3.

Relationships

• related: The current balance of power in open models (Nathan Lambert, September 2026) — Qwen as the source of China's Hugging Face download lead and of academic adoption (September 2026)
• related: DeepSeek, China and the US Are Running Different AI Races, Open-Source AI / Open-Weight Models, Distillation, Fast-Follow Problem, AI Sovereignty
• related: Khanna letter to Alibaba on pacing the frontier (September 16, 2026) — addressee of Rep. Ro Khanna's September 2026 letter on frontier pacing; see Coordinated slowdown proposals compared
• contradicts: the "Alibaba tilting closed" framing in Open-Source AI / Open-Weight Models and (Source: epochai.substack.com) — Qwen3's Apache 2.0 235B release is counter-evidence. The closed, API-only Qwen3.7-Max release (May 2026) aligns with the "tilting closed" framing at the flagship tier, even as the open Qwen3 family persists in lower tiers; Alibaba's posture is bifurcated (open mid-tier, closed flagship).