🤖 AI Weekly Recap (Week 33)

Tiny but Real Time: At 100M parameters it runs comfortably on ordinary CPUs, targeting edge devices, mobile apps, and local installs where latency, cost, and privacy matter.

Clone Any Voice Locally: You can clone a voice from any audio sample when you run the model on your own machine, so nothing ever leaves your computer.

One Command Setup: Run uvx pocket-tts generate and it handles dependencies on the fly, or pocket-tts serve to spin up a local web interface at localhost:8000.

90,000+ Free Skills: The Hub auto-refreshes twice daily and includes contributions from NVIDIA's skills repo, added to trusted taps in v0.16.0, alongside community registries.

Cross-Agent Standard: Skills are just SKILL.md files, and the format is compatible across Hermes, Claude Code, Cursor, Codex, and other agents.

Install and Publish in One Line: It connects to official skills, GitHub repos, the skills.sh marketplace, and ClawHub through one unified interface, and you can publish your own set as a tap with hermes skills tap add owner/repo, no server or registry signup required.

Biggest Qwen Ever, By a Lot: It is a huge jump from Qwen3-Max at 1 trillion parameters in September 2025 and Qwen3.5 at 397 billion in February 2026, and developer Shuai Bai called it the team's first multimodal model above 1 trillion parameters, processing text, images, video, and documents.

The Number Nobody Has: It is a sparse MoE design like the rest of the Max tier, but the active-parameter count has not been disclosed, and without it the 2.4T headline says very little about actual serving cost.

Cheap to Test Right Now: A preview is available through Alibaba's Token Plan subscription, its Qoder coding platform, and QoderWork productivity suite at 10% of standard pricing.

Built for Long Horizon Work: It targets complex engineering across large repositories, planning changes, writing code, and validating results, with a set of async background agents that stay alive for the whole session instead of spawning per task.

Your Files Are Never Touched: Multiple agents work in parallel in their own isolated worktrees, and as Zuckerberg put it, your working copy is never touched. A local append-only event log records every model call, tool run, approval, and edit, which Meta calls replay-exact and restart-safe.

Cheap, With an Asterisk: It installs with one shell command on macOS and Linux, runs approvals and an OS sandbox by default, and costs $1.25/$4.25 per million tokens, or $0.10/$0.20 if you let Meta train on your code. Independent harnesses are less flattering than Meta's own charts, so test it on your own repo.

Point and Edit: A magic wand tool changes only the region you point at, segmentation selects precise areas, and background removal exports any subject with transparency for use in other software.

Five References at Once: Multi-reference editing accepts up to five input images in a single generation, removing the manual compositing step that used to mean stitching sources together by hand. A smart resize tool automatically fills the frame when you change aspect ratio.

Number 2 in the World: On the Arena Image Edit and Text-to-Image leaderboards it ranks second globally in both, behind OpenAI's GPT-Image-2. Worth noting those are xAI's own figures pulled from a public leaderboard, and API access is still listed as coming soon.

Storage as Memory: Only 16 of the model's 896 experts per layer fire for any given token, so the always-on trunk stays in RAM while the sleeping experts stream from disk and get multiplied straight out of their packed 4-bit form. 93% of the 1.56 TB checkpoint never enters memory, read from NVMe with O_DIRECT.

Same Answer at Any Budget: The identical model runs in 8 GB and in 224 GB and produces byte-identical output at every budget in between. More RAM only buys speed, never accuracy.

Fully Open, Genuinely Slow: Released under Apache 2.0, and honest about itself: at 32 seconds per token on a laptop and 10 seconds on a high-RAM workstation, it is a proof of concept, demonstrating that the hardware barrier to frontier models is a storage problem, not a compute problem.