Daily digest
10 items · ~10 min · Week 2026-W31
Must-read (4)
Anthropic CEO Clarifies: Company Does Not Oppose Open-Weight Models
AnthropicOn July 27, 2026, Anthropic CEO Dario Amodei published an official statement saying Anthropic has never advocated for banning open-weight AI models and considers them a public good when they lack dangerous capabilities, while voicing concern about frontier-capable systems reaching the wrong hands.
Moonshot Kimi K3 Open-Weight Release: 2.8T Parameter MoE Model on Hugging Face
Moonshot AIMoonshot AI released Kimi K3 on Hugging Face on July 26-27, 2026: a 2.8-trillion-parameter sparse MoE model with a 1-million-token context window and native text/image/video multimodal input. At 1.4TB with MXFP4 quantization, it is the largest open-weight model released to date, priced at $3/$15 per million tokens via API; demand reportedly forced Moonshot to temporarily suspend new API subscriptions.
SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD
SLAISLAI T-Rex is an end-to-end optimization framework for training trillion-parameter MoE models on Ascend NPU hardware, achieving 34.22% Model FLOPs Utilization — a 2.93x improvement over baseline — while maintaining large-scale training stability.
Self Gradient Forcing: Native Long Video Extrapolation
Self Gradient Forcing (SGF) is a two-pass training strategy for autoregressive video diffusion that restores missing supervision in the context-gradient path, enabling extrapolation to multi-minute videos from only 5-second training windows with improved identity preservation and temporal stability.
Worth knowing (5)
Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
Microsoft ResearchMage-Flow is a 4-billion-parameter generative system combining a lightweight latent tokenizer (Mage-VAE) with a native-resolution multimodal diffusion transformer, generating images from 512 to 2048 pixels at any aspect ratio with 2.5x faster training throughput.
Claude Code v2.1.219 Sets Claude Opus 5 as Default Opus Model
AnthropicClaude Code v2.1.219 (July 24) defaults to Claude Opus 5 (1M-token context) for complex agentic tasks, adds nested subagent forwarding for depth-2+ subagents, and a sandbox.network.strictAllowlist option for network access control. v2.1.220 (July 25) followed with bug fixes.
GitHub Copilot Adds Claude Opus 5 for Agentic Coding Tasks
Anthropic / GitHubClaude Opus 5 became available in GitHub Copilot (July 24) for Pro+, Max, Business, and Enterprise users across VS Code, JetBrains, and Xcode, targeting multi-step coding workflows with autonomous code changes and regression verification.
OpenAI Codex Adds ChatGPT Voice on Desktop for Hands-Free Agentic Coding
OpenAIChatGPT Voice extended to the Mac and Windows desktop app (July 23, build 26.715) with direct integration into Codex, using OpenAI's full-duplex GPT-Live voice model so a spoken instruction can be handed to multiple autonomous coding agents.
MCP 2026-07-28 Specification Released with Stateless Architecture and Extensions Framework
The Model Context Protocol reached its 2026-07-28 final specification with a stateless core architecture, an extensions framework, and OAuth/OIDC-aligned authorization. Beta SDKs are live for Python, TypeScript, Go, and C#.
For reference (1)
OpenCode v1.18.5 Fixes Multi-LLM Compatibility Issues
OpenCode v1.18.5 (July 24) improves Claude adaptive-thinking response handling, fixes an OpenAI response-phase bug that broke conversations, stabilizes Mistral prompt caching, and preserves grep symlink paths in search results.