Reflection AI unveils Beam, an open-weight coding and agent model it says matches Z.ai’s GLM-5.2 reasoning and coding performance with a fraction of the compute, with Apache 2.0-licensed weights due this month.
DeepSeek releases DeepGEMM Ascend, using DeepJIT runtime kernel compilation to avoid CUDA builds at installation and a fused Mega MoE kernel that overlaps NVLink transfers with tensor-core computation.
Reflection unveils Beam, a 501-billion-parameter Mixture-of-Experts reasoning model with 23 billion active parameters, and says it will release Apache 2.0-licensed weights and developer tools later this month.
Cognition adds persistent Memory and daily “Dreaming” to Devin, enabling the coding agent to retain each user’s preferences, corrections and project lessons in a Git-backed Markdown repository across sessions, while open-sourcing its agent-memory standard.
Gatewai launches the beta of Framefields, an Apache-2.0 Node.js package that renders code-first video directly via WebGPU, Dawn, Metal or Vulkan instead of headless Chromium, while offering on-device ONNX vision tools for segmentation, pose tracking and person mattes.
Companies are accelerating moves to open-source and sovereign AI after Anthropic’s June 30-day chat-retention mandate for Fable and Mythos raises privacy concerns, prompting Booz Allen to restrict Fable use before new zero-data-retention options arrive Sept. 1.
Meta open-sources Muse AI gadget code and SDKs while opening a waitlist for 5,000 Home Link devices shipping this month, enabling developers to run the agent on ESP32 boards or Raspberry Pis to control lights, TVs and printers at their own risk.
0xpili launches a GitHub skill that steers LLMs toward ASD-STE100 Simplified Technical English, bundling 53 rules, an 869-word approved vocabulary, rewrite examples, and a Python 3 checker that flags passive voice, long sentences, contractions, and unapproved words.
Runway launches Runway Ads, an AI agent that creates, edits, publishes and tracks ad campaigns, alongside Praxis-1, an open-weight robotics world-action model, as CEO Cristóbal Valenzuela projects fully world-model-powered physical-AI deployments in 12 to 18 months.
Strands launches the open-source Decider 2B model for low-latency local AI-agent decisions, reporting roughly 115 ms median latency on an Nvidia RTX 3090 and a top JevBench ranking among strictly 2B-parameter models.