Chinese AI labs lead open-source push with DeepSeek and Alibaba releases

DeepSeek open-sourced its agent harness (Cordis meta-framework) under MIT license and released V4 Pro as open-weight on Hugging Face under MIT, a strategic shift after raising API pricing up to 1,100% for V4 models. This move signals a bid to capture developer ecosystem and compete with Nvidia's open-source strategy. The open-weight release allows enterprises to run the model on their own infrastructure, potentially disrupting the closed-model market. A new thesis on DeepSeek reveals Liang Wenfeng's singular vision: betting on learning as the path to AGI, with open-source as a strategic bet on future dominance. New pricing data shows V4-pro is now 2x cheaper per task than V4-Flash, indicating aggressive cost optimization. Alibaba released Qwen3.8 as open-weight under Apache 2.0, featuring a massive 2.4T-parameter model and a 27B variant that runs on consumer GPUs. The 27B model outperforms Qwen3.7-Plus in coding and office tasks, with 262K native context and improved agent capabilities. It requires ~54GB BF16, 27GB FP8, or 14-16GB 4-bit, making it viable on 24GB consumer cards. Alibaba's open-source models have hit 3B downloads, surpassing Meta and Google, reinforcing their growing dominance in the open-source ecosystem. New data: Qwen 3.8 27B performs 2x faster on Atomic Agent vs Hermes/Prime, adding practical deployment data for agent platform selection.