DeepSeek Leaks MODEL1: New Flagship AI Shocks The Industry

DeepSeek is rumored to be preparing a new flagship AI model, 'MODEL1', with significant architectural changes, while Zhipoo AI released GLM 4.7 Flash for efficient coding and reasoning, Japanese researchers developed an AI to model emotions based on bodily signals, and Nouse Coder 14B demonstrated advanced competitive programming capabilities through reinforcement learning.

• DeepSeek is poised to release 'MODEL1', a new flagship AI model, with substantial architectural changes indicated by code leaks.

• Zhipoo AI's GLM 4.7 Flash delivers a powerful, locally deployable AI for advanced coding and reasoning with extensive context support.

• New research in Japan enables AI to model human emotions from bodily signals, while Nouse Coder 14B excels in competitive programming via direct code execution training.

Evidence suggests DeepSeek is developing a new flagship AI model, 'MODEL1', which may represent a structural redesign rather than a minor iteration, indicated by changes in KV cache, sparsity handling, and FP8 decoding support, potentially integrating recent research like MHC and Ngram. Zhipoo AI's new GLM 4.7 Flash offers a powerful yet deployable solution for coding and reasoning with long context, targeting users unable to host massive models. Furthermore, Japanese researchers at NAIST created an emotion computing AI based on the theory of constructed emotion, aligning with human self-reports and having potential applications in mental health, while Nouse Coder 14B has emerged as a top-performing AI in competitive programming by learning through code execution and explicit reward/punishment.

DeepSeek's Next AI Model

• 00:00:21 Developers discovered evidence of DeepSeek's potential next flagship AI model, referred to as 'MODEL1', through extensive code updates on GitHub. This 'MODEL1' identifier appears frequently alongside references to DeepSeek v3.2, suggesting it is not a minor version update but a structurally new model, possibly DeepSeek V4. The timing for this unveiling is speculated to be mid-February, around the Lunar New Year period, with a major focus on improved coding ability.

Architectural Changes in MODEL1

• 00:01:53 Analysis of the code updates reveals significant technical differences between 'MODEL1' and v3.2, indicating deep architectural changes rather than superficial improvements. These include redesigns in KV cache layout, which impacts memory efficiency and speed in long contexts, as well as differences in sparsity handling, suggesting compute efficiency tricks. Additionally, the explicit support for FP8 decoding points to an engineering focus on efficiency at scale, aligning with DeepSeek's recent research into new training methods like Modified Hierarchical Connections (MHC) and the Ngram memory module.

Zhipoo AI's GLM 4.7 Flash

• 00:05:17 Zhipoo AI released GLM 4.7 Flash, a 31 billion parameter Mixture of Experts (MoE) model designed for strong reasoning and coding capabilities while being realistic enough to run locally without massive GPU clusters. This model supports 128,000 tokens of context and is positioned as a lightweight, free-tier option compared to their full GLM 4.7 model, competing effectively in benchmarks for math, reasoning, and coding agent tests. It also supports various inference methods like VLLM and SGLANG, with an ecosystem for fine-tuning and quantization.

Emotion Computing AI and Nouse Coder

• 00:08:52 Japanese researchers developed an AI framework for emotion computation based on the theory of constructed emotion, using a multi-layered, multi-modal latent Dirichlet allocation (MMLDA) model. This AI learns emotion categories from patterns in physiological responses, visual input, and language, achieving 75% agreement with human self-reported emotions, suggesting applications in emotion-aware robots and mental health support. Separately, Nouse Coder 14B, an AI for competitive programming, was released by News Research, trained with reinforcement learning where code is executed and rewarded only for correct solutions, achieving a 7.08 percentage point boost over its base model on LiveCodeBench v6, demonstrating mastery of real-world test cases.