Kimi K3, DeepSeek V4 Pro, GLM-5.2 Models Compared

Three Chinese research labs have released large, open-weight Mixture-of-Experts (MoE) models that now lead the open leaderboard. Moonshot AI's Kimi K3, DeepSeek V4 Pro, and Zhipu AI's GLM-5.2 all feature million-token context windows and are designed for long-horizon coding and agent tasks. This comparison evaluates them based on measured capability, licensing terms, and the practical cost of serving them.

Kimi K3, released on July 16, 2026, is a 2.8-trillion-parameter Stable LatentMoE model. It activates 16 out of 896 experts per token, with Moonshot AI not disclosing the exact active parameter count. Kimi K3 includes native vision and video capabilities, a 1 million-token context window, and always-on reasoning, positioning itself as the first open 3-trillion-parameter class model. DeepSeek V4 Pro, released on April 24, 2026, is a 1.6-trillion-parameter MoE model with 49 billion active parameters, utilizing 384 routed experts plus one shared expert. It also offers a 1 million-token context window with a maximum output of 384,000 tokens. A smaller variant, DeepSeek V4 Flash, has 284 billion total parameters and 13 billion active parameters for more cost-effective workloads.

GLM-5.2, released by Zhipu AI on June 13, 2026, is a 744-billion-parameter MoE model with approximately 40 billion active parameters and a 1 million-token context window. It provides High and Max reasoning modes and is available via API access. While GLM-5.2 is the smallest of the three by total parameters, it held the top spot in the open-weight field before Kimi K3's release. The article notes that vendor-reported benchmark scores for these models use different methodologies, making direct comparison challenging.

The comparison focuses on three key decision-making axes for AI teams: capability, license, and serving cost. Kimi K3 and DeepSeek V4 Pro are described as 'trillion-parameter' models, with Kimi K3 at 2.8T and DeepSeek V4 Pro at 1.6T. GLM-5.2, with 744 billion total parameters, is the smallest but remains competitive due to its performance and features. The models' licensing terms and the associated infrastructure costs for deployment are critical factors for adoption and practical use.