Hello from The AI Night,
Today in AI:
Anthropic Launches Voice Mode In Claude Code Terminal
Alibaba’s Qwen Releases 4 New Qwen 3.5 Small Models
Anthropic Just Made Memory Available On Their Free Plan
Image Source: Thariq tweet
Here's the deal: Anthropic is rolling out voice mode in Claude Code, starting with roughly 5% of users today and expanding over the coming weeks. Users with access will see a notification on the welcome screen.
The Breakdown:
Hold spacebar to talk, release to send. Transcript streams directly at your cursor position.
Voice input works inline with typed text. You can type part of a prompt, speak the middle section, then continue typing. It does not overwrite existing text.
Voice transcription tokens do not count against rate limits and there is no extra cost.
Available on Pro, Max, Team and Enterprise plans on a rolling basis.
Toggle it on with the /voice command.
The bigger picture: This lowers friction for developers who think faster than they type, especially when describing complex logic or debugging flows mid-prompt. Inline mixing of voice and text is a practical workflow detail that separates this from simple dictation tools.
Image Source: Qwen tweet
Here's the deal: Alibaba's Qwen team released four compact models under the Qwen 3.5 series: 0.8B, 2B, 4B and 9B parameters. All share the same Qwen 3.5 architecture with native multimodal support, improved design and scaled reinforcement learning. Base model weights are also available.
The Breakdown:
All four models are natively multimodal, not text-only retrofits. This is uncommon at the sub-10B scale.
0.8B and 2B target edge devices where speed and size constraints dominate.
4B is positioned as a lightweight base for multimodal agents.
9B is described as closing the gap with much larger models, though no specific benchmarks were shared in the announcement.
Weights are available on both Hugging Face and ModelScope.
The bigger picture: Small, capable, open-weight multimodal models lower the barrier for on-device AI, embedded agents and resource constrained research. If the 4B and 9B perform as implied, they give builders practical alternatives to closed APIs for lightweight production use cases.
Image Source: Claude tweet
Here's the deal: Anthropic expanded Claude's memory feature to its free plan and launched a memory import tool that lets users bring stored context from ChatGPT, Gemini or other AI providers into Claude.
The Breakdown:
Memory was previously limited to paid plans (Pro, Max, Team, Enterprise) since its launch in late 2025.
The import process works through a copy-paste prompt, users paste Anthropic's provided prompt into their current chatbot, then paste the exported data into Claude's memory settings.
Users retain full control. They can view, edit, pause or permanently delete all stored memories.
Claude is currently the #1 free app on the US App Store. Free users are up 60% since the start of 2026 and paid subscribers have doubled.
ChatGPT already offers memory on its free tier, making this a competitive catch-up move.
The bigger picture: This lowers the switching cost for users locked into other AI ecosystems. By removing the paywall on memory and simplifying data portability, Anthropic is converting its recent consumer momentum into long-term retention.
What else you need to know:
Wispr Flow launched a public dictation speed challenge after 3.5 million people watched its initial test with five users, offering a Porsche GT3 RS to anyone who can type faster than its voice input.
Google's NotebookLM fully rolled out its Slide Revisions feature on mobile, letting users edit and revise AI-generated presentation slides directly from their phones without needing a desktop.
Vercel Labs released an Electron skill for its agent-browser framework, enabling coding agents to control desktop apps like Discord, Figma, Notion, Spotify, and VS Code programmatically.
Cognition previewed SWE-1.6, trained with 100x more compute than SWE-1.5 on the same base model, scoring 11% higher on SWE-Bench Pro while maintaining 950 tok/s inference speed.
Google's Stitch now lets users manually edit text and images or use an agent to make targeted updates directly on screens, adding precision control alongside its AI-driven design generation.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.