China OSS surge + Kimi K3 + DeepSeek V4 GA + Qwen 3.8 Max + Treasury sanctions + distillation accusations + ban threat
Key Questions
What is Kimi K3 and when were its weights released?
Kimi K3 is a 2.8T-parameter model from Moonshot AI that topped Arena.ai benchmarks. Its open weights were released on July 27, 2026, with day-0 hosting on platforms like Hugging Face and Together.
How does Kimi K3 compare to closed models like Fable 5?
Kimi K3 outperformed GPT-5.6 on several benchmarks with a 6x cost advantage and showed high agreement on DeepSWE tasks. It achieved #1 ranking on Arena.ai while requiring massive resources like 1.49TB storage.
What US policy actions target Chinese open-weight models?
The US Treasury has threatened sanctions on Chinese AI models, and the White House accused Moonshot of distilling Fable. Experts question the timeline feasibility, and industry groups including Nvidia and Microsoft have pushed back against broad restrictions.
Why are distillation claims against Kimi K3 disputed?
White House claims cite specific platform evidence, but experts note the Fable release date makes distillation implausible. Counterpoints highlight that post-training acceleration and synthetic data blur the line with true weight copying.
What hardware is needed to run Kimi K3 locally?
Kimi K3 requires 1.49TB storage and 64+ accelerators, making local deployment impractical for most users. GGUF conversions and extreme quantization have been attempted but remain experimental and not fully loadable.
How do Chinese open-weight models affect token usage and revenue?
They account for 29% of token volume on Vercel's gateway but only 4% of spend, showing strong adoption driven by cost. Open-source models now power 33% of usage while generating just 4% of revenue.
What industry coalition supports open-weight AI?
Nvidia, Microsoft, Meta, Mistral, and Hugging Face signed a joint letter defending open weights and opposing restrictions. The Linux Foundation and AI Tinkerers have also endorsed the movement citing security and innovation benefits.
What is the status of Qwen 3.8 and DeepSeek V4?
Alibaba's Qwen 3.8 claims 2.4T parameters and second place after Fable 5, with weights expected soon. DeepSeek V4 GA is available under MIT license with 80.6% SWE-bench performance at 1/57th the cost of closed alternatives.
China OSS surge continues with Kimi K3 (2.8T params, 1M context) beating Fable 5, weights released under permissive license with $20M revenue cap. A new practical guide shows building Kimi K3 in pure C to run on just 8GB RAM via disk streaming, pushing local deployment boundaries. DeepSeek V4 GA (MIT, 80.6% SWE-bench, 1/57th cost). DeepSeek V4 Flash now available in LM Studio with US-hosted cloud option and local deployment guide (requires 156GB RAM). DeepSeek V4-Flash confirmed as cheapest major model to run (~$0.03 per benchmark). Qwen 3.8 Max announced with open weights coming next week (2.4T params, 1M context, multimodal, agentic) – a strategic pivot from Alibaba. Qwen3.8-Max debuts on Arena.AI; self-hosting requires 1.2TB VRAM at 4-bit, 95B active params. Qwen3.8-27B variant more practical for local deployment (confirmed to run on 17GB RAM/VRAM). A dedicated intelligence, performance & price analysis of Qwen3.8 Max provides concrete data for community evaluation. MiniMax H3 (video generation) also released. US Treasury threatens sanctions; White House accuses Moonshot of distilling Fable. Policy fight intensifies with Lambert warning of possible US ban in 6 months. Deltafin claims to run Kimi K3 in just 64GB RAM. Qwen-UI-Agent released. Mozilla's CTO highlights economic gap. Moonshot's Kimi uses 20k Nvidia chip cluster from Alibaba. Anthropic clarified no ban, Amodei targets Chinese AI. New articles: open-closed gap is nuanced—task evolution and RL environments are the real moat, but Chinese labs are closing in. Kimi K3 is still data-center only; practical alternatives exist. Latest weekly roundup adds DeepSeek Seedance 2.5 (video generation) and Minimax H3 as new open-weight releases from China. Community excitement around Qwen3.8-27B as a practical local model. Best open-source coding LLMs list highlights DeepSeek V4 Pro and Qwen3.8-27B. New article 'The death zone' highlights 80% of US startups using Chinese open-source models, reinforcing competitive pressure. A critical take on DeepSeek Flash claims it is below Grok 4.5 and Muse Spark. Latest: Chinese AI models narrow gap with US frontier labs by 4-8 months, with 2-10x cost advantage, driving market shift. GLM-5.2 also narrowing gap. A tweet shows Qwen3.8-Max can be prompted for object detection with boxes, achieving 60-80% mAP, demonstrating practical multimodal capability. Qwen3.8-Max pricing at $2/M tokens and autonomous coding for 16 days reported. New today: African developers using Chinese open-weight models for local tools, reinforcing global impact. Article on Kimi K3 and Arizona chips adds cynical framing of open weights as marketing ploy.