AI generation and Claude Code updates #104
Today's Letter
MiniMax-H3 Turbo LoRA, 4-step audio-video generation
- MiniMax-H3 Turbo LoRA generates video with synchronized stereo audio in as few as 4 sampling steps
- The model card reports a reduction from roughly 20 steps to 4, or about a 5× sampling speedup
- v4 step 600 is the recommended checkpoint for static and small-motion content, with improved fine detail
- Six to eight steps generally improve quality; four steps can produce trailing artifacts during fast motion
- The LoRA uses strength 1.0 with the simple scheduler and supports text-to-video and image-to-video workflows
- ComfyUI integration is provided through ComfyUI-MiniMax-H3-Turbo, while standalone generation is supported by generate.py
- BF16 weights are approximately 744 MB and still require the MiniMax-H3 base model, VAEs, and text encoder
Source: huggingface.co
More: news.hada.io · simonwillison.net
Claude Code, cross-session messaging introduced
- Claude Code can send text messages between independent sessions on the same machine
- The feature requires Claude Code v2.1.224 or later and runs on macOS and Linux
- ListAgents discovers reachable sessions, while SendMessage delivers messages by session name
- Messages can carry findings, breaking changes, decisions, and status updates without sharing files or conversation history
- Permission boundaries remain separate for each session, and inbound messages can be delivered, held, or refused
- Cross-machine sessions can reply through Remote Control, but cannot initiate the exchange
- Resume, agent teams, agent view, Remote Control, and channels remain separate features for other coordination patterns
Source: code.claude.com
More: macrumors.com
Jocoletter curates AI, software, and product trends for developers and builders.
#Anthropic #HuggingFace