In September 2026, five Chinese AI coding models—Kimi K3, Qwen3.8 Max, DeepSeek 4.1 Flash, GLM-5.3 and MiniMax M3—compete with Claude and GPT on code writing, all offering downloadable weights and licenses permitting commercial use. According to the article, users can now install on their own server a model solving three quarters of SWE-bench Verified tasks without paying a subscription. In early September, Artificial Analysis wrote that open source had caught up with closed models, with most important open weights carrying Chinese names from Alibaba, DeepSeek, Moonshot AI, Zhipu and MiniMax. BenchLM's overall index, averaging reasoning, tool-calling and long tasks, ranks Kimi K3 (Moonshot AI) first at 74.4 with open weights, strong in frontend and long agentic tasks; Qwen3.8 Max (Alibaba) at 73.2 under Apache 2.0 for huge context and local use; GLM-5.3 (Zhipu/Z.ai) at 66.9 under MIT for repository agents and terminal; DeepSeek 4.1 Flash (DeepSeek) at 63.7 under MIT for value and SWE-bench; and MiniMax M3 (MiniMax) at 61.3 with open weights for tool-calling and 1 million context. The gap between first