xAI power delays, DeepSeek V4 Flash update, TensorRT-LLM bugs, EU AI Act enforcement

🚀 Models & Releases

DeepSeek V4 Flash 0731 released: sparse MoE model with 13B active params (284B total), optimized for coding, reasoning, and agent workflows. Available on Hugging Face (deepseek-ai/DeepSeek-V4-Flash-0731) and OpenRouter (deepseek-v4-flash-0731). Tags include 8-bit, fp8, and region:us. (Hugging Face, OpenRouter)
Inkling Small by Thinking Machines Lab: open-weight multimodal MoE model (12B active/276B total), positioned as a smaller, efficient alternative. (OpenRouter)

(No significant research updates in this brief.)

xAI power plant delays: SpaceX won’t remove unpermitted turbines for xAI’s Colossus data centers for another year, potentially impacting infrastructure timelines. (TechCrunch)

llama.cpp b10215: Vulkan driver version check added for Windows Intel GPUs to prevent crashes (fixed from driver 32.0.101.8860). Supports Xe2+ GPUs. (GitHub)
TensorRT-LLM v1.3.0rc23: Known issues include hangs/crashes with DeepSeek-V4-Pro (GB300 disagg setups), DeepSeek-R1 (MPI worker exits), and Qwen MoE multi-LoRA. (GitHub)
cline SDK v0.0.69: Ollama’s response-start timeout extended to 5 minutes (from 30s) to accommodate large model cold-loads. (GitHub)
crewAI 1.15.10: Adds skill usage event collection. (GitHub)
pydantic-ai v2.22.0: Defaults Gemini VALIDATED tool mode on supported models. (GitHub)

OpenAI Python v2.52.0: Adds content provenance checks and honors Retry-After delays up to 2 minutes. (GitHub)

EU AI Act enforcement begins: European Commission tools now operational for compliance. (Euractiv)
US AI regulation momentum: Forbes outlines five drivers for impending federal AI policy. (Forbes)
Minnesota nudify ban upheld: Judge rejects xAI’s request to block state law targeting AI-powered image manipulation apps. (TechCrunch)

xAI’s power infrastructure delays could disrupt Colossus data center timelines—monitor for cascading impacts on model training/deployment.
DeepSeek V4 Flash 0731 offers a cost-efficient MoE alternative for coding/reasoning workflows; evaluate against existing providers.
TensorRT-LLM v1.3.0rc23’s bugs (DeepSeek/V4-Pro hangs, Qwen MoE crashes) may require workarounds for GB300 users.