🤖 AI Weekly Recap (Week 32)

13+ Platforms, One Command: Covers Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, Douyin and more, with built-in diagnostics (agent-reach doctor) that detect and repair broken setups automatically.

Actually Zero Cost: Every backend is a free open-source tool (yt-dlp, twitter-cli, rdt-cli, Jina Reader, Exa), so nothing requires a paid key. The only optional expense is a residential proxy at roughly $1/month if you are running from a server. For comparison, the official X API runs around $215/month for moderate use.

Works With Everything You Already Use: Any agent that can run shell commands can use it, including Claude Code, Cursor, Windsurf, Codex and OpenClaw, and it also ships as an MCP stdio server for framework-level integration.

Selective Section Painting: You can edit a specific part of a track, or expand a short melody into a full song, without regenerating from scratch. That means correcting just a verse or just a chorus.

Vocals That Stop Sounding Synthetic: Google cites more expression and emotion, more realistic singing, and improved pronunciation, alongside richer melodic structures and lyrics with better prompt adherence and structural awareness. Pronunciation has always been the giveaway on AI vocals, so this is the most testable claim in the set.

Real Producer Controls: Tracks run from 30 seconds to 3 minutes, with fine control over tempo and duration across vocals, drums, bass and other elements. It is free for all Flow Music users and everything is SynthID watermarked.

30 Seconds in One Take: Up to 30 seconds of high-quality audio-video generated in a single pass, with multi-round extensions on top. That is double Seedance 2.0's 15-second ceiling, and the model arranges multiple logically connected shots so a story runs through setup, development, turning point and resolution.

50 Multimodal References: It accepts up to 50 reference inputs including images, audio clips, 3D white models and style references, up from 12 in the previous version. The split is roughly 30 images, 10 videos and 10 audio clips per generation.

Native 4K, Not Upscaled: Generation happens at 4K rather than being upscaled from lower resolution, with 10-bit color depth for grading headroom. ByteDance also claims 20 percent better prompt adherence, and audio is co-processed in the same latent space as visuals for native sync.

Whisper-Large Swap Fixes the Lips: Version 1.5 replaces the old Wav2Vec2 audio encoder with Whisper-Large, producing noticeably smoother and more natural lip dynamics. Wav2Vec2 was trained mostly on English, while Whisper-Large-v3 covers 680,000 hours across 99 languages, so accented and multilingual speech finally maps cleanly.

Three Native Tasks, One Model: Audio-Text-to-Video, Audio-Text-Image-to-Video, and audio-driven video continuation, with single-stream and multi-stream audio, plus multi-person scenes and stable identity across long shots, singing, and fast speech.

Built to Actually Run: 8-step distilled inference with optional INT8 quantization for lower VRAM, at 480P and 720P output. Fair warning from hands-on testing: it still wanted a 40GB A800 and roughly 44 seconds of GPU time per second of finished video.

284B Total, Only 13B Active: A sparse Mixture-of-Experts with 284 billion total parameters and 13 billion active per token, a 1M-token context window, and 384K max output. The router picks top-6 from 256 routed experts plus one shared expert on every token.

The Small Model Beat the Big One: The changelog is explicit that it "was only re-post-trained," yet agent benchmark scores now far exceed V4-Pro-Preview, the much larger model in its own family, at roughly a third of the output price. DeepSeek's card puts 0731 at 82.7 on Terminal-Bench 2.1 versus the April preview at 61.8.

Absurdly Cheap and Fully Open: $0.14 per million input tokens and $0.28 per million output, against category medians of $0.58 and $2.20. Weights went to Hugging Face under MIT. One caveat worth knowing: the model is verbose, burning 210M tokens on the Artificial Analysis eval suite against a 62M median, so real cost per task lands above the sticker price.

It Has Its Own Computer: Most agents live inside a text box and hand code back for you to copy, paste and debug. By giving the AI its own machine, ByteDance removed the handoff. Stateful memory and a persistent filesystem mean it remembers your writing style, project structures and preferences across separate sessions.

Lead Agent Plus Spawned Sub-Agents: A lead agent acts as project manager and decomposes a prompt like "research the top 10 AI startups in 2026 and build me a presentation" into logical sub-tasks, then spawns sub-agents in parallel rather than grinding through one linear chain of thought. Each sub-agent gets its own scoped context, tools and termination conditions.

Many features: It works with any OpenAI-compatible provider, and ships with Docker sandboxing, MCP servers, and built-in skills that cover research, report generation, slide creation, web pages, and image and video generation, loaded progressively so the context window stays lean.