Alibaba unveils biggest AI model as DeepSeek sets new low-cost benchmark
Alibaba introduced Qwen3.8-Max, its largest AI model, with 2.4 trillion parameters and a mixture-of-experts design using 95 billion at a time. The model is due next week and Alibaba said it completed a software-engineering task in 16 days. Shares rose 7% in Hong Kong. Research firm Artificial Analysis said DeepSeek’s V4-Flash costs $0.14 per million input tokens and $0.28 per million output, about 3 cents per test versus higher peers.
How this was made
The 30-second read
Why it matters
Alibaba’s disclosed model specs and stated next-week release create a near-term catalyst for sentiment and potential developer interest. DeepSeek’s reported per-token pricing is a competitive benchmark that can influence how investors think about AI inference margins and pricing power.
Market read
Traders may treat this as a competitive benchmark reset for AI inference costs and a near-term catalyst for Alibaba into its next-week model release.
What to watch
Developer adoption, enterprise integration, and compute supply constraints (chips, inference capacity) may matter more than headline token pricing or parameter scale.
Background
The article frames a fast-moving China AI race focused on open-weight models and affordability, contrasting Alibaba’s Qwen3.8-Max with DeepSeek’s V4-Flash cost benchmarks.
Ticker impact
Alibaba unveiled Qwen3.8-Max with 2.4T parameters and said it will be released next week, lifting its Hong Kong shares 7%.
Shares may remain bid into the next-week release as leaderboard performance and developer adoption expectations build.
The article provides a same-day share reaction (up 7% in Hong Kong) plus concrete model specs (2.4T parameters, MoE with 95B active) and a stated release timeline.
Market effects
Reinforces a shift toward open-weight, cost-optimized inference (MoE, low per-token pricing), which can reset competitive expectations across AI model providers.
Supports China tech sentiment via a high-profile model launch and leaderboard momentum.
Benchmark comparisons (vs Claude and Kimi) highlight global inference-cost competition, potentially affecting cross-border AI capex and pricing assumptions.
Counterpoint
Leaderboard ranking and parameter counts may not translate into durable revenue; pricing benchmarks can be promotional and may not reflect sustained unit economics.
Key entities
- companyAlibaba
Introduced Qwen3.8-Max, its largest AI model to date, and reported a same-day share jump in Hong Kong.
- productQwen3.8-Max
Alibaba’s open-weight AI model with 2.4 trillion parameters and a mixture-of-experts design using 95B active parameters.
- companyDeepSeek
Released V4-Flash and is cited for extremely low inference pricing versus other well-known models.
- productV4-Flash
DeepSeek model with reported $0.14 per million input tokens and $0.28 per million output tokens.
- research_firmArtificial Analysis
Provided the cost comparisons and benchmark estimates cited in the article.
