Underdog Woof launches a 31.7MB text-to-speech model that runs in browsers and offline. The two-voice version is 40.6MB and achieves 1.42% word error rate on hard sentences, 25% fewer errors than Fish Audio Drama 3.
Key points
Size: one-voice model 31.7MB, two-voice model 40.6MB.
Accuracy: two-voice Woof 1.42% word error rate on hard sentences, 25% fewer errors than Fish Audio Drama 3.
Sources
- Original post
- Underdog Woof - Linked page
Underdog Woof 31.7MB TTS model released
Underdog Woof is a 31.7MB text-to-speech model that runs in the browser via WebGPU and can generate voices offline. It claims to beat larger TTS models on Word Error Rate and is released under CC BY-NC 4.0 with open weights.
Key points
Size: 31.7MB, runs in browser with WebGPU
License: free and open weight under Underdog CC BY-NC 4.0
Sources
- Original post
- underdog.ai - Linked in the post
OCRia-1 launched as best general purpose document parser
OCRia-1 is announced as the best general purpose model for document parsing, handling complex text layouts, tables, diagrams, and charts. It powers all modes within LlamaParse and is benchmarked on ParseBench. Two OCR approaches are described: fast/cheap free OSS tools integrated in liteparse, and accurate VLM-based OCR for critical workflows.
Key points
OCRia-1 launched as a general purpose document parsing model.
Liteparse integrates with five open-weight OCR tools, including MiniCPM-OCR.
Sources
- Original post
- github.com - Linked in the post
- llamaindex.ai - Linked in the post
- parsebench.ai - Linked in the post
Opus 5.5 builds iOS app in minutes, no Xcode, no Mac, no App Store
Opus 5.5 builds a high-quality, bespoke iOS app in minutes. The post contrasts this with hours needed for Xcode, App Store Connect and TestFlight setup and days for review.
Key points
Opus 5.5 can build a high-quality, bespoke iOS app in minutes
It takes hours to get Xcode + App Store Connect + TestFlight to work properly, and days to get it reviewed
Sources
- Original post
- ios.vibecodeapp.com - Linked in the post
Vvoyer uses an agent to investigate errors in ClickHouse
Vvoyer uses an agent to investigate errors in ClickHouse data, illustrating an agent-led observability approach. Vercel shows what that can look like. Training on ClickStack is available.
Key points
Vvoyer uses an agent to investigate errors in ClickHouse data.
Vercel demonstrates the agent-led observability approach.
Sources
- Original post
- clickhou.se - Linked in the post
Claude memory dreaming announced by Anthropic
Claude uses a “dreaming” process to organize learned information, correct mistakes, and improve future retrieval. Anthropic shared this detail in an engineering blog post.
Key points
Engineering Blog post: “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.
Anthropic says Claude does something similar to the memory-consolidation approaches discussed for Letta_AI and LangChain.
Sources
- Original post
- wired.com - Linked in the post
ScreensDesign MCP server launches, Full Pro required
ScreensDesign MCP server launches, Full Pro required, lets AI assistants search app screens, flows, revenue signals and review analytics. Requires the Full Pro plan. Supports multiple AI models.
Key points
Hosted MCP server lets AI assistant search mobile app screens, full flows in order, revenue and install signals, and App Store review analytics
Works with ChatGPT, Claude, Claude Code, Cursor, Codex, OpenClaw and Windsurf