Open-Weight Model Competition: China vs. Meta
Key Questions
What are the key specifications of Moonshot's Kimi K3 model?
Kimi K3 is a 2.8T-parameter open-weight model with a 1M context window, released on Hugging Face. It joins other Chinese models like GLM-5.2 in gaining traction in the US market.
Why are Chinese AI models gaining US adoption?
They offer lower costs and 'good enough' performance that challenges US closed-source leaders. Hosting requires about 1.5TB VRAM under mxfp4 quantization, with fine-tuning remaining a noted challenge.
What investment implications arise from Moonshot's Kimi K3 release?
Moonshot's $50B IPO valuation adds significant investment interest in open-weight Chinese models. The release highlights shifting competitive dynamics between US and Chinese AI ecosystems.
Moonshot released Kimi K3 (2.8T, open-weight, IPO at $50B), Alibaba Qwen3.8-Max (2.4T MoE, open-source next week, $2/M tokens, text-only), MiniMax H3 (video+audio). Meta responded with Muse Spark 1.2 (strongest US open-weight) and Muse Glimmer (30B local agent). DeepSeek released V4-Flash-Vision (multimodal open-weight) and V4 Pro under MIT license. GLM-5.3 achieves 50% coding improvement without scaling, winning KingBench 3. Mystery model OX Alpha (likely Zhipu GLM-5.x) corrected to 58.4% DeepSWE, ~10% behind GLM-5.3. GLM-5.5 claims top open-weights coding model. DeepSeek Flash usage surging 10-100x in enterprises confirms open-source cost advantage. Open models consistently beat closed frontier models at pass@k for k>1. DeepSeek raising prices 2-3x on V4 Flash. Alibaba AI models hit 3B downloads. AMD enables Qwen3.8 27B on Ryzen AI and Radeon GPUs. MOSS-VL (11B open vision-language model) released for real-time video understanding. Nvidia moving to build own open-weight models via Poolside deal, directly competing with customers and validating open-weight trend.