DeepSeek vs LiteLLM

Open-weight AI models gain production traction

The article discusses the shift in the AI industry from closed to open-weight models, citing that in August 2026, open models processed 56% of tokens on Vercel's AI Gateway, up from 7% in December 2025, and about 60% of US-originating token consumption on OpenRouter. While closed models still lead at the frontier, open weights are becoming part of production infrastructure. The piece introduces the concept of 'open escape velocity', where open AI develops independent sources of demand and infrastructure, reducing dependence on any single model company. However, the content is largely paywalled, and the full analysis and data are not accessible.

• Open-weight models processed 56% of all tokens on Vercel's AI Gateway in August 2026, up from 7% in December 2025.
• OpenRouter reported open models accounting for roughly 60% of US-originating token consumption in the same period.
• DeepSeek demonstrated that open-weight models can compete on cost, capability, and deployability.

Global Times: Italian Scientist Praises China's AI, Tech Innovation Success

Global Times interviewed Italian professor Francesco Faiola, who says China's innovation success, including the C919 airliner, DeepSeek, Unitree Robotics, and Moonshot AI's Kimi K3, is now real global news noticed even by Italian retirees. Faiola sees China's openness to international collaboration as a sign of confidence and highlights the need for joint funding, harmonized regulations, and talent circulation between China and Europe. He dismisses 'China Shock 2.0' as lacking evidence, emphasizing that innovations like DeepSeek create new markets.

• Chinese AI startup Moonshot AI released its Kimi K3 large language model ahead of the 2026 World AI Conference.
• Kimi K3 is described as the world's largest open-source model by parameter count to date.
• Chinese President Xi Jinping stressed accelerating high-level self-reliance in science and technology for the 15th Five-Year Plan period (2026-2030).

DeepSeek V4.1 Flash Beats Larger Model, Proving Ilya Right

Ilya Sutskever left OpenAI to found SSI and predicted the end of scaling. DeepSeek's new V4.1 Flash model, with a third of the parameters of its flagship V4-Pro, outperforms it on coding benchmarks (74.2 vs 62.7) while activating fewer parameters per token. Starting September 14, all V4-Pro requests route to V4.1 Flash. This validates the shift from model size to efficiency, marking a turning point in AI development.

• DeepSeek V4.1 Flash has 552 billion parameters, activating 8-16 billion per token, while V4-Pro has 1.6 trillion parameters, activating 49 billion.
• V4.1 Flash scores 74.2 vs V4-Pro's 62.7 on DeepSeek's coding tests.
• All V4-Pro requests are rerouted to V4.1 Flash starting September 14.

Auto Router: 45% Lower Cost on 25 SWE-bench Tasks

We solved 23 of 25 SWE-bench Verified tasks with LiteLLM's experimental capability router for $11.15, compared with $20.27 using Opus 5.