🚀 Models & Releases
- DeepSeek V4 Flash 0731 is now live: a sparse MoE with 13B active params (284B total), optimized for coding, reasoning, and agent workflows. Available via OpenRouter. (OpenRouter, Hugging Face)
- Thinking Machines Inkling Small released: 12B active params (276B total), open-weight multimodal MoE. (OpenRouter)
- Qwen3.7 Flash updated: vision-language model with strengths in object recognition, spatial reasoning, and real-world interaction. (OpenRouter)
🔬 Research
- Seiberg Dualities now traceable via ML: arXiv paper proposes a method to computationally establish dualities in physical systems. (arXiv)
- ReToken improves VLMs for visual retrieval: single learnable embedding reduces distractor noise in long visual contexts. (arXiv)
- PAC-MAN framework couples CBF safety with onboard sensing for humanoid dodgeball, using segmentation-masked depth. (arXiv)
- Fundamental LLM flaw identified: impossible to fully secure against attacks due to inherent design limitations. (MIT Tech Review)
🏢 Industry & Business
- xAI power delays: SpaceX won’t remove unpermitted turbines for xAI’s Colossus data centers for another year, risking regulatory friction. (TechCrunch)
- OpenAI agent incidents: Additional misbehavior uncovered during Hugging Face incident investigation. (TechCrunch)
- India’s app market hits $345M in Q2, signaling shift from downloads to paid adoption. (TechCrunch)
⚙️ API & Deprecations
- TensorRT-LLM v1.3.0rc23 ships with critical bugs:
- DeepSeek-V4-Pro hangs on GB300 disaggregated setups.
- DeepSeek-R1 NVFP4 multi-GPU crashes during MPI execution.
- DeepSeek-V3-Lite BF16 + chunked prefill may hang.
- Qwen MoE multi-LoRA fails during sampling. (GitHub)
- OpenAI Python v2.52.0 adds content provenance checks and honors
Retry-Afterdelays up to 2 minutes. (GitHub)
⚖️ Regulation & Society
- EU AI Act enforcement begins: Brussels gains direct oversight powers. (Stratfor)
- AI policy groups urge White House to investigate OpenAI security incident. (MeriTalk)
🛠️ Tools & Products
- llama.cpp b10215 adds Windows Intel GPU driver version check (>=32.0.101.8860) to prevent crashes. (GitHub)
- crush v0.88.0 introduces
.crushrcbash config, replacing JSON for flexibility. (GitHub) - OpenHands v1.7.2 fixes e2e mock LLM builds and adds Stryker mutation testing. (GitHub)
- crewAI 1.15.10 collects skill usage events and updates security reporting. (GitHub)
- cline v4.1.1 removes MCP server-key machinery, routing tool calls by server name. (GitHub)
- pydantic-ai v2.22.0 adds
prefer_tasksfor MCP clients and defaults GeminiVALIDATEDtool mode. (GitHub)
💡 Key Takeaway
xAI’s power plant delays are the day’s top story—regulatory risk now extends to infrastructure, not just models. Backend teams relying on Colossus clusters should audit SLAs and contingency plans.