Cur8 — Tuesday, September 22, 2026

Jev, a new type of LLM called a "decision model," outputs probabilistic decisions instead of text, priced at $0.042/million tokens, cheaper than GPT-5 Nano. It excels in classification tasks.

Hugging Face integrates llama.cpp quantization into Transformers, enabling local AI on Apple Silicon with GGUF models. Q4_K_M variants reduce Qwen3.5-4B size to 2.74 GB.

Jun Kim ensures stability and long-term development of oMLX.

Xiaomi's Mimo-V2.6 Pro outperforms Kim K3 and Quen 3.8 Max on benchmarks, with 1 million token context.

GPT-6 Astra broke the 2005-resistant Enigma message MVUEH (10 July 1941) on 15 September 2026, using the crib "ROSENOW" and discovering a unique key.

OpenAI proposes global AI standards focusing on coordinated evaluation, reporting, and governance to enhance safety and transparency. The initiative aims to align industry practices and improve accountability in AI development.

OpenAI collaborates with an independent advisory group to oversee and communicate emerging AI research in mathematics and AI. The group aims to ensure rigorous evaluation and transparent dissemination of results.

GPT-6 Astra enables Higgsfield AI to deploy new video features rapidly, accelerating ad creation for small businesses with enhanced creative tools.

OpenAI Academy expands with new learning paths targeting employees, developers, leaders, educators, and students to build and demonstrate practical AI skills.

Grok 4.6, with 500K token context and xhigh reasoning, is available on Amazon Bedrock from August 18, 2026, supporting long-running agents and coding. It scores 61 on AA Intelligence Index and 1753 on GDPVal-AA v2.

Grok 4.7 outperforms Grok 4.6 on CursorBench 4.0 and GDPval, and tops LatchBio’s biosafety benchmark at 62.4%. Priced from $2/million input tokens, it offers enhanced safeguards and red-team capabilities for cybersecurity.

Jev is a fast decision-making AI, up to 200x faster than chatbots, using parallel sampling and RLCD training. It offloads simple decisions from complex systems, reducing costs and token usage, but lacks a research paper and builds on older concepts.

Anthropic suspended Fable 5 and Mythos 5 due to US export controls, causing industry realignment. Chinese models like GLM 5.2 and new architectures like OpenRouter’s Fusion API gained traction as alternatives.

Transformers, introduced in 2017, use self-attention to process sequences and power models like GPT and Llama. GPT-2 (small) has 12 blocks and 12 attention heads, converting text to embeddings via tokenization, embedding, and positional encoding.