SourceSCMP Technology•Stalecollected in 3m
Alibaba Unveils New Qwen Models and Custom AI Chips
Alibaba is building a full-stack 'AI factory' to challenge global leaders in model training and inference infrastructure
30-Second TL;DR
What Changed
Introduction of new Qwen model iterations for enhanced training and inference.
Why It Matters
Alibaba's move to vertically integrate hardware and software signals a major push to dominate the Chinese AI infrastructure market. This could significantly lower costs for developers relying on Alibaba Cloud for large-scale model deployment.
What To Do Next
Review the Alibaba Cloud documentation for the latest Qwen model APIs to evaluate their performance against current open-source alternatives.
Who should care:Developers & AI Engineers
Key Points
- •Introduction of new Qwen model iterations for enhanced training and inference.
- •Development of custom AI chips to support large-scale model workloads.
- •Strategic shift toward 'AI factory' model, focusing on revenue generation through compute-intensive services.
- •Integration of cloud infrastructure with autonomous agent capabilities.
Key numbers30%50%
Deep Insight
Background and context from public sources — not the original article. 15 sources cited.
Enhanced Key Takeaways
- •Alibaba's AI-related product revenue has reached an annualized run rate of CNY 35.8 billion (approximately USD 5.2 billion) and currently accounts for 30% of its Cloud Intelligence Group's external revenue, with expectations to exceed 50% within a year.
- •The newly introduced Zhenwu M890 AI chip, developed by Alibaba's T-Head subsidiary, delivers three times the performance of its predecessor, the Zhenwu 810E, and is specifically engineered to handle the high memory and communication demands of autonomous AI agent workloads.
- •Alibaba has committed to an investment exceeding RMB 380 billion (approximately USD 53 billion) over three years in cloud and AI infrastructure, indicating a significant capital expenditure to support its 'AI factory' ambitions.
- •The Qwen model family includes both open-source (e.g., Qwen 3.5-7B, MIT-licensed) and proprietary variants, with some open-weight models demonstrating superior performance in benchmarks like MMLU compared to competitors such as GPT-4o-mini, at significantly lower inference costs.
- •Alibaba's 'AI factory' strategy is driven by the rapid expansion of AI agent workloads, encompassing training, inference, and orchestration, with current demand for compute capacity reportedly outstripping available supply.
Competitor Analysis
MMLU Score
- Alibaba Qwen 3.5-7B (Alibaba Cloud API)
- 74.2%
- OpenAI GPT-4o-mini
- 72.9%
- Anthropic Claude 3.5 Haiku
- N/A
- Google Gemma 3-9B
- N/A
Input Price (per 1M tokens)
- Alibaba Qwen 3.5-7B (Alibaba Cloud API)
- $0.008
- OpenAI GPT-4o-mini
- $0.15
- Anthropic Claude 3.5 Haiku
- $0.08
- Google Gemma 3-9B
- $0.03
Output Price (per 1M tokens)
- Alibaba Qwen 3.5-7B (Alibaba Cloud API)
- $0.01 (Together AI)
- OpenAI GPT-4o-mini
- N/A
- Anthropic Claude 3.5 Haiku
- N/A
- Google Gemma 3-9B
- N/A
License
- Alibaba Qwen 3.5-7B (Alibaba Cloud API)
- MIT-licensed (for open-weight models)
- OpenAI GPT-4o-mini
- Proprietary
- Anthropic Claude 3.5 Haiku
- Proprietary
- Google Gemma 3-9B
- Proprietary
On-device Inference
- Alibaba Qwen 3.5-7B (Alibaba Cloud API)
- Yes (0.5B model on iPhone 15 Pro at 40 tokens/sec)
- OpenAI GPT-4o-mini
- No (API-only)
- Anthropic Claude 3.5 Haiku
- No (API-only)
- Google Gemma 3-9B
- N/A
| Feature/Metric | Alibaba Qwen 3.5-7B (Alibaba Cloud API) | OpenAI GPT-4o-mini | Anthropic Claude 3.5 Haiku | Google Gemma 3-9B |
|---|---|---|---|---|
| MMLU Score | 74.2% | 72.9% | N/A | N/A |
| Input Price (per 1M tokens) | $0.008 | $0.15 | $0.08 | $0.03 |
| Output Price (per 1M tokens) | $0.01 (Together AI) | N/A | N/A | N/A |
| License | MIT-licensed (for open-weight models) | Proprietary | Proprietary | Proprietary |
| On-device Inference | Yes (0.5B model on iPhone 15 Pro at 40 tokens/sec) | No (API-only) | No (API-only) | N/A |
Technical Deep Dive
- Qwen Model Architecture: The Qwen family includes both dense transformer models (like Qwen 3.5, optimized for edge hardware) and Mixture-of-Expert (MoE) architectures (like Qwen 2.5-Max).
- Qwen Training Data: Qwen 2.5-Max was pretrained on over 20 trillion tokens, encompassing multi-lingual textual data and domain-specific corpora.
- Qwen Multimodality: Models like Qwen-Omni are end-to-end multimodal, processing text, images, audio, and video, and delivering real-time streaming responses.
- Qwen Thinking Modes: Qwen3 models feature hybrid 'Thinking' and 'Non-Thinking' modes, allowing flexible control over reasoning performance, speed, and costs.
- Zhenwu M890 AI Chip: This custom chip features 144 gigabytes of GPU memory, an upgrade from its predecessor's 96 gigabytes, enabling it to process significantly larger data for complex AI agent workloads.
- Hanguang 800 AI Chip (Predecessor): Alibaba's first AI inference chip, launched in 2019, was built on a 12-nm process with 17 billion transistors. It achieved a peak performance of 78,563 images per second (IPS) on ResNet-50 inference tests.
Future ImplicationsAI analysis grounded in cited sources
Alibaba's AI-related product revenue will become the primary driver of its Cloud Intelligence Group's external revenue.
Management expects AI-related product revenue to exceed 50% of external cloud revenue in about a year, pivoting the business more toward AI compute and agent services.
Alibaba will continue to rapidly advance its custom AI chip technology with a clear roadmap for future generations.
The company has outlined a multi-year chip roadmap, planning to launch the V900 in Q3 2027 and the J900 in Q3 2028, each expected to deliver significant performance gains.
The Qwen model family will solidify its position as a foundational 'operating system' for AI development, particularly in the open-source community.
Alibaba remains committed to open-sourcing Qwen models and aims to shape it into the 'operating system of the AI era,' empowering developers globally.
Timeline
2009-09
Alibaba Cloud officially established.
2017-Late
DAMO Academy, Alibaba's research institute, launched.
2018-09
T-Head, Alibaba's semiconductor division, spun out of DAMO Academy.
2019-09
Alibaba unveils Hanguang 800, its first AI inference chip.
2023-04
Alibaba launches beta of Tongyi Qianwen (Qwen) large language model.
2026-05
Alibaba unveils Zhenwu M890 AI chip and Qwen 3.7-Max model.
- 2009-09Alibaba Cloud officially established.
- 2017-LateDAMO Academy, Alibaba's research institute, launched.
- 2018-09T-Head, Alibaba's semiconductor division, spun out of DAMO Academy.
- 2019-09Alibaba unveils Hanguang 800, its first AI inference chip.
- 2023-04Alibaba launches beta of Tongyi Qianwen (Qwen) large language model.
- 2026-05Alibaba unveils Zhenwu M890 AI chip and Qwen 3.7-Max model.
Sources (15)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
1nationalcioreview.comvertexaisearch.cloud.google.com2yicaiglobal.comvertexaisearch.cloud.google.com3economictimes.comvertexaisearch.cloud.google.com4morningstar.comvertexaisearch.cloud.google.com5moomoo.comvertexaisearch.cloud.google.com6eweek.comvertexaisearch.cloud.google.com7mindstudio.aivertexaisearch.cloud.google.com8udit.covertexaisearch.cloud.google.com9deepinfra.comvertexaisearch.cloud.google.com10medium.comvertexaisearch.cloud.google.com11alibabacloud.comvertexaisearch.cloud.google.com12medium.comvertexaisearch.cloud.google.com13chinadaily.com.cnvertexaisearch.cloud.google.com14eetimes.comvertexaisearch.cloud.google.com15alibabacloud.comvertexaisearch.cloud.google.com
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.