Zhipu AI's GLM-OCR reaches 94.62% on OmniDocBench v1.5 despite using only 0.9B parameters. I dug into its layout parsing, vertical text handling, and math recognition.
An open-source TTS model distilled from the ZipVoice architecture into four inference steps, delivering voice cloning with 1 GB of VRAM and 150x real-time speed. It also compares itself with the other TTS models covered on this blog.
Added --autoConnect option to Chrome DevTools MCP server, allowing coding agents to connect directly to an existing browser session. A deep dive into the background of the MCP vs CLI debate, browser operations with OpenClaw, and the risks of authenticated session delegation.
Agentic pipeline, which combines ReACT loop and searcher, achieved 1st place in ViDoRe v3 and 2nd place in BRIGHT. We established versatility for multiple domains using the same pipeline.
Anthropic has GA’d a 1M‑token context window. No surcharge for long context; image/PDF per‑request limit raised from 100 to 600. Achieved a frontier‑model best score on MRCR v2.
Anthropic's new multi-agent code review feature for Claude Code, plus the design split between subagents and orchestration. Also covers the major frameworks and where Codex fits in.
OpenAI acquired AI security evaluation platform Promptfoo, and Microsoft announced that Anthropic's Claude Cowork would be integrated into Microsoft 365 Copilot. The structure of the enterprise AI market is starting to change.
Sarvam AI released 30B and 105B models trained entirely in India—from pretraining through RL—featuring support for 22 constitutionally recognized Indian languages and inference optimizations.
Andrej Karpathy released Autoresearch, a system where an AI agent autonomously runs machine-learning experiments on a GPU and tries 100 variants overnight. The article breaks down the mechanism and design so even readers with zero ML background can follow.
Anthropic found 22 CVEs in Firefox's JS engine with Claude, while GitHub Security Lab reported more than 80 vulnerabilities in apps built on the OSS framework Taskflow Agent.
WAN 2.2 image-to-video on Windows + RTX 4060 8GB VRAM in ComfyUI. The 5B fp8 model failed three times; the 14B Rapid distilled model with --lowvram offloading produced a 2-second clip in 111 seconds — vs 82 minutes on M1 Max 64GB. Working setup and what to avoid.
