🤖 AI Weekly Recap (Week 21)

Moodboards Done Right: Upload up to four reference images (or scale a single moodboard up to 250 images) and K2 blends their combined styles, colors, textures, mood, and composition, into every output, holding the aesthetic consistently across generations.

Aesthetic-First Architecture: Krea's team poured as much engineering into the style transfer system as into the foundation model itself, training on curated editorial photography, fine art, and concept work rather than generic image scrapes.

Fast and Flexible: Generates outputs in 15 seconds or less, with adjustable style strength sliders and the ability to mix multiple references. A serious creative tool, not a one-size-fits-all generator.

Full-Duplex Architecture: The model breaks conversation into 200-millisecond "micro-turns," letting it perceive tone shifts, facial expressions, and interruptions while still generating its own response. No more walkie-talkie style turn-taking.

Faster Than the Giants: At 0.40-second latency, TML-Interaction-Small beats comparable models from OpenAI and Google, including Gemini-3.1-flash-live, on the FD-bench benchmark.

Encoder-Free Early Fusion: A 276B-parameter model handles real-time chat while a second model runs deeper reasoning and web search in parallel. Raw audio and video stream directly into the transformer instead of going through separate translation encoders.

One Screen for Everything: Open with claude agents and see every session at a glance: running, waiting for input, completed, failed, or idle. Spacebar peeks at the latest turn, Enter attaches to the full transcript.

Background-First Workflow: Send any session into the background with /bg, launch new background jobs with claude --bg "", and reply inline without losing context across dozens of parallel runs.

Safe Parallelism: Background sessions automatically write to isolated git worktrees, so 10+ agents can edit the same repo without stomping on each other. Available now for Pro, Max, Team, Enterprise, and API plans.

40+ Built-In Tools: Orchestrates frontier LLMs (Claude Opus 4.7, GPT-5.5 Pro, Gemini 3.1 Pro and more) alongside image and video models under one chat. No tool-hopping between ten different platforms.

Three Layers of Memory: Stores brand guidelines, past campaign history, and creative preferences in long-term memory so the agent improves with every run. "Make another like the third one" actually works across sessions.

Runs 24/7 From Anywhere: Schedule tasks daily, weekly, or one-off, and access entirely via browser or Telegram. Trigger a full ad campaign render from your phone while you're on the bus.

Efficient at Scale: A 295B-parameter MoE architecture that activates only 21B parameters per token (top-8 of 192 experts), delivering competitive performance against trillion-parameter models at a fraction of the compute cost.

Agentic Powerhouse: Scores 74.4% on SWE-bench Verified, 67.1% on BrowseComp, and 54.4% on Terminal-Bench 2.0, and has reliably powered agentic workflows up to 495 steps long in production.

Three Reasoning Modes: Pick between no_think, think_low, and think_high per request to trade off latency and depth. Open weights available on GitHub for full self-hosting and fine-tuning.

True Background Capture: Twelve Claude Code lifecycle hooks fire automatically (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, and more). No manual add() calls. The agent works, the hook fires, the memory lands.

Works Everywhere: One universal MCP JSON snippet plugs into Claude Code, Claude Desktop, Cursor, VS Code, OpenCode, Codex, Gemini CLI, Cline, Roo Code, and Windsurf. Basically any agent that speaks MCP.

Real Retrieval Benchmarks: 95.2% R@5 on LongMemEval-S, 92% fewer tokens than equivalent context-stuffing approaches, and zero external databases. Runs as a single Node process with state on disk. Apache-2.0 licensed.