Tech • IA • Crypto
China Just Dropped a Real Threat to Anthropic
8/10TL;DR
A rapid श्रृ of Chinese AI releases, led by Alibaba’s Qwen 3.8 Max, signals intensifying global competition driven by scale, efficiency, and a deep domestic talent pipeline.
KEY POINTS
Alibaba launches Qwen 3.8 Max
Alibaba unveiled Qwen 3.8 Max, a 2.4 trillion-parameter multimodal model positioned among the largest globally. It ranks as the top Chinese text model on Arena AI and second worldwide on visual benchmarks, just behind an Anthropic system. The model handles text, images, and video with up to 1 million tokens of context, enabling large-scale document and code processing.
Efficiency through mixture-of-experts
Despite its scale, Qwen activates only 95 billion parameters per task using a mixture-of-experts design. This reduces compute cost and latency while maintaining high performance. The system reportedly completed a full software engineering project in 16 days, highlighting its applied capabilities.
Coordinated wave of releases
On July 31, three major Chinese labs—MiniMax, ByteDance, and DeepSeek—simultaneously launched new systems. The clustering of releases underscores intense domestic competition and rapid iteration cycles across video generation and AI infrastructure.
MiniMax H3 reshapes video workflows
MiniMax’s H3 introduces a unified multimodal video system capable of generation, editing, and composition in a single pipeline. It supports 2K output and up to 15 seconds of audiovisual content per run, while integrating text, image, audio, and video inputs. The model leads video editing benchmarks and is priced at roughly 0.8 yuan per second, significantly below competitors.
Commercial implications for advertising
H3 enables advanced motion transfer, allowing users to swap characters, products, and environments while preserving original motion. This capability streamlines ad production by turning raw assets—product images, brand elements, and reference clips—into finished marketing videos with minimal manual editing.
ByteDance extends narrative video generation
ByteDance’s Seedance 2.5 doubles single-pass video length to 30 seconds and focuses on narrative consistency. The model maintains characters, scenes, and style across sequences, enabling multi-minute video creation through continuous generation. It also supports up to 30 images, 10 videos, and 10 audio inputs for detailed creative control.
Improved realism and control in video AI
Seedance 2.5 reduces common artifacts such as inconsistent characters and artificial textures. It allows timeline-level editing, including precise control over actions, camera angles, and scene changes, signaling a shift from clip generation to full production workflows.
DeepSeek prioritizes cost efficiency
DeepSeek V4 Flash emphasizes affordability rather than scale, achieving performance gains through post-training improvements rather than larger models. It is priced at $0.14 per million input tokens and $0.28 per million output tokens, with an average task cost of $0.03—over 100 times cheaper than leading competitors like Claude Fable 5.
Competitive performance at low cost
On a composite intelligence index, V4 Flash scores 50/100, matching mid-tier global models while dramatically undercutting them in cost. The model is optimized for agent-based tasks such as coding, tool use, and workflow automation, reflecting a shift toward practical deployment efficiency.
Talent pipeline challenges US assumptions
Research from Stanford HAI and the Hoover Institution shows 84.5% of DeepSeek-affiliated researchers are currently based in China, with 53.5% trained entirely domestically. Among researchers with US experience, 70% returned to China, challenging the notion of sustained Western talent dominance.
Rising domestic capability in China
The findings highlight a growing “systems capability” in China, where large teams and institutional pipelines produce frontier AI systems at scale. Models such as DeepSeek R1 and Moonshot Kimi K3 were reportedly developed at a fraction of the cost of comparable Western systems.
CONCLUSION
A surge of releases combining massive scale, lower costs, and strong domestic talent signals a तेजी escalation in the global AI race, with Chinese labs closing performance gaps while redefining efficiency and deployment economics.