Model Launches & Open Weights
Kimi K3 launches. Chinese company Moonshot AI released Kimi K3, an open-weight model that matches frontier performance on some benchmarks, beating Claude Fable 5 on SpreadsheetBench. The release stirred debate about China's open-source AI strategy and contributed to a dip in chip stocks.
Deepseek V4 incoming. Deepseek is teasing the imminent release of V4, promising low-cost API access and performance comparable to Kimi K3 and Fable, continuing its aggressive price-performance optimization.
German-English MoE model. The German SooFi team released Soofi S 30B-A3B, an open-source Mixture-of-Experts hybrid Mamba-Transformer model for German and English, targeting efficient bilingual deployment.
openPangu-2.0-Flash added. Support for openPangu-2.0-Flash, a 92B-A6B model with 512K context and novel features like MLA-latent cache, has been added to ik_llama.cpp.
Ternary models go sub-2-bit. PrismML's Bonsai line uses true ternary weights (~1.7 bits/weight) by converting existing models like Qwen3. An 8B version fine-tuned on metal shows competitive everyday performance at roughly 2 GB.
Agent Infrastructure & Tools
Pinecone Nexus for agent knowledge. Pinecone launched Nexus Engine, a knowledge engine that compiles enterprise context into a structured layer AI agents can query directly, reducing per-query retrieval loops and improving accuracy.
Catch cache invalidation easily. Cache-hunter is a new tool that helps harness builders detect cache invalidation in LLM calls by visualizing session stability, revealing unstable prompts and tools across many popular harnesses.
Corporate Moves & Policy
Anthropic backtracks on Fable 5 limits. Competition from GPT-5.6 Sol and Kimi K3 forced Anthropic to reverse its plan to remove Claude Fable 5 from subscriptions. Max and Team Premium plans will include it at 50% of reduced limits, while Pro users get a one-time $100 API credit.
Pentagon's AI-first fleet strategy. The Department of the Navy approved a strategy to weaponize data and AI, prioritizing rapid deployment over perfect alignment to build an 'AI-first' fleet through a five-stage adaptation cycle.
China launches parallel AI order. China announced the World Artificial Intelligence Cooperation Organization (WIKO) with 29 Global South nations, excluding Western countries, signaling a push for a parallel AI governance structure.
FastFlowLM joins AMD. FastFlowLM, known for efficient inference solutions, has been acquired by AMD, aiming to advance AI inference on AMD hardware.
Eggers confronts OpenAI over writing. Author Dave Eggers told OpenAI staff that ChatGPT is 'silencing an entire generation' by enabling students to avoid learning to write, calling the impact on educators 'catastrophic'.
Research & Techniques
KV cache grafting improves routing. Researchers demonstrated byte-exact KV cache grafting on frozen Gemma 4, boosting a routing system's accuracy from 76.7% to 90.0% on AIME 2025 by reusing verified knowledge as KV state.
Multi-mode reasoning in LLMs. An in-depth article explains how reinforcement learning with verifiable rewards can train LLMs to support low-, medium-, and high-effort reasoning modes, similar to GPT-5.6's settings.
Open-weight cyber gap closing fast. The UK's AISI found that open-weight models now match the cyber capabilities of closed frontier models from four months ago at a fraction of the cost, with safeguards easily bypassed.
That's everything for this day – about a 5-minute read.