JetBrains released Mellum2 Thinking, a 12B Mixture-of-Experts model with 2.5B active parameters, built for low-latency reasoning inside agentic coding pipelines.
FinRL splits into a research sandbox and FinRL-X, a production-ready trading system with modular architecture, LLM support, and live broker integration
agentic-stack is an open-source portable .agent/ folder that gives Claude Code, Cursor, Windsurf, and five other harnesses a shared memory and skills layer that survives tool switches.
xAI's terminal coding agent Grok Build opens its beta to all SuperGrok and X Premium+ subscribers, bringing Plan Mode, parallel subagents, Imagine, and CI-ready headless scripting.
Kuaishou's new 30B mixture-of-experts vision-language model ships DeepSeek Sparse Attention to push 256K context and hour-long video understanding into open source.
Qwen3.7-Max now caches repeated prompt prefixes automatically, cutting input costs by up to 90% with zero code changes — and offers explicit control for power users who need guaranteed hits.
Comfy-Org repackages NVIDIA's VAE-free PixelDiT model for ComfyUI, making the CVPR 2026 pixel-space diffusion transformer plug-and-play for local image generation.
OpenBMB's MiniCPM5-1B packs a 128K context window, hybrid reasoning, and native tool-calling into a model that fits on a phone — and beats every competitor in its weight class.