1 day ago
DeepSeek Narrows AI Performance Gap With US Rivals to 3%
A new report says Chinese AI models are getting closer in performance to leading US models.
The gap is now 3%, compared with 9% in May and 15% earlier this year.
DeepSeek V4.1 Flash ranked sixth on a test called LiveBench.
Its score was 81.1, while Anthropic’s model scored 83.4.
Still, only three of the 15 models on the ranking were Chinese.
Analyst Robert Lea said Chinese companies are improving models to work with the hardware they can access.
He said this may reduce the effect of US technology and chip restrictions.
But China’s AI companies face strong competition and low prices, and Lea expects the industry may not become profitable until 2030.
Bloomberg Intelligence said Chinese AI models trail US models by 3% on benchmark scores, down from 9% in May and 15% earlier this year.
DeepSeek V4.1 Flash ranked sixth globally on LiveBench and was the highest-ranked Chinese model.
DeepSeek scored 81.1 on LiveBench, compared with 83.4 for Anthropic’s model.
Only three of the 15 models ranked on LiveBench were Chinese, according to Bloomberg Intelligence.
Analyst Robert Lea said China’s AI industry may remain unprofitable until 2030 amid competition among more than 1,100 language models.
- Who
- Chinese AI companies, including DeepSeek, and US AI rivals; Bloomberg Intelligence analyst Robert Lea commented on the trend.
- What
- The performance gap between Chinese and US AI models narrowed to 3%, while concerns about competition and profitability remain.
- Where
- The comparison concerns Chinese and US AI models and their results on LiveBench.
- When
- The report was published on Monday; it cites DeepSeek V4.1 Flash’s September launch and benchmark results for the month.
- Why
- Bloomberg Intelligence attributed China's progress to growing AI expertise and researchers optimizing models for available hardware.
Evidence of narrowing competition
Limits and unresolved challenges
China’s progress and US technology restrictions
Evidence of narrowing competition
Lea said Chinese firms are improving models by optimizing them for available hardware, potentially weakening the impact of US restrictions on technology and NVIDIA chips.
Limits and unresolved challenges
The report also notes that Chinese models remain a minority on LiveBench: only three of the 15 ranked models were Chinese.
Performance gains and commercial viability
Evidence of narrowing competition
DeepSeek V4.1 Flash ranked sixth globally, with a score close to Anthropic’s model.
Limits and unresolved challenges
Lea said intense competition among more than 1,100 models and a token price war could keep China’s AI industry unprofitable until 2030.
Key facts
- Reported performance gap
- Chinese AI models trailed US models by 3% on benchmark scores.
- Previous gap
- The gap was 9% in May and 15% earlier in the year.
- DeepSeek model
- DeepSeek V4.1 Flash launched in September and ranked sixth worldwide on LiveBench.
- LiveBench scores
- DeepSeek V4.1 Flash scored 81.1; Anthropic’s model scored 83.4.
- Chinese models ranked
- Three of the 15 LiveBench-ranked models were Chinese.
- Market competition
- The Chinese market had more than 1,100 language models, according to the report.
- Profitability outlook
- Robert Lea said China's AI industry might remain unprofitable until 2030.
Quotes
Robert Lea
Bloomberg Intelligence senior analyst
“casts doubt on the long-term stability of US technological supremacy in artificial intelligence.”
livemint.com
Sources
Related news
China Questions Whether AI Slowdowns Could Strengthen US Lead
Beijing Rejects Anthropic CEO’s Call to Curb China’s AI
Nvidia-Backed Reflection Unveils Open-Weight AI Model Beam
Trump Faces Difficult Choices in Contest With China Over AI
US-China AI Dialogue Highlights Divergent Approaches to Frontier Risks
AI Slowdown Debate Expands Into a Global Race
Reasoning Models Can Help AI Bypass Safety Guardrails
China Emphasizes Control as US AI Firms Warn of Risks
Delhi Government Opens AI Training Centre for Officers
