Technical Developments in llama.cpp

Work has resumed on implementing Multi-Token Prediction (MTP) support for the Qwen Flash Next model within the llama.cpp framework. According to reports on Reddit, developers have restarted work on the MTP implementation, which is currently categorized as a work-in-progress (WIP).

Users seeking to test this functionality can access GGUF quantizations for Qwen3.8-Flash-Next via Hugging Face. While a Pull Request (PR) has been submitted to the ggml-org/llama.cpp repository to facilitate this, the implementation remains in an experimental stage.

Community discussions on Reddit suggest that previous attempts to implement MTP and direct reading from the Position Embedding Layer (PLE) faced delays. Some users noted that while direct PLE reading may not be the final implementation path—with developers potentially favoring mmap+prefetch methods—current workarounds for direct PLE reading have reportedly increased performance by approximately 50% to 80% depending on the specific scenario. However, users cautioned that the current performance of certain experimental models remains suboptimal.

Alibaba and the Expansion of the Qwen Ecosystem

ALIBABA GROUP HOLDING continues to expand its presence in the open-source AI sector. The company has released Qwen3, described as its first hybrid reasoning AI model, and has previously released Qwen3-Coder, designed for software development and complex coding workflows.

Data from Hugging Face indicates that Alibaba’s Qwen models have reached significant scale, with reports suggesting 700 million downloads. Furthermore, Alibaba has released specialized medical AI models capable of detecting cancer and nearly 150 other conditions. In the robotics sector, the company’s RynnBrain has reportedly set multiple records in AI-robotics benchmarks.

Global AI Market Dynamics and Chinese Model Adoption

The landscape of AI development shows significant activity from Chinese firms despite ongoing U.S. restrictions on talent and chip technology. Recent data indicates that Chinese AI models have captured 41% of global Hugging Face downloads and have accounted for up to 46% of developer use in the United States.

Key developments in the sector include:

  • Model Performance: DeepSeek-V3.1 has been optimized for Chinese-made chips, and DeepSeek V4-Flash has reportedly topped global token consumption rankings on OpenRouter as of July 2026.
  • Infrastructure: Baidu has reported the successful illumination of a cluster comprising 30,000 self-developed P800 Kunlun chips. Additionally, China has launched its first AI data center powered by 100% green electricity.
  • Research and Talent: Reports suggest that 38% of the world’s top AI researchers were educated in China.

While U.S. efforts have included bans on U.S. persons working at Chinese chip firms and restrictions on assisting the Chinese AI sector, Chinese companies continue to advance in specialized fields, including solar observation, geological mapping, and drug screening. For instance, Tsinghua University researchers have developed DrugCLIP, a platform intended to accelerate drug screening.