Z.ai's FP8 GLM-5.2 halves the memory to self-host the top open coder
The MIT coder that tops GPT-5.5 at a sixth the cost now fits one 8-GPU node — plus Z.ai's open character animator and Copilot's multi-repo CLI agents.
Copy markdown
Half the memory, a sixth the cost
Z.ai's official FP8 build of GLM-5.2 — 744B parameters (~40B active), MIT-licensed, 1M-token context — drops the weights to about 744GB, half the BF16 footprint, so you can serve it from a single 8xH200 node instead of a two-box cluster. It posts 62.1 on SWE-bench Pro (past GPT-5.5's 58.6) and ~74% on FrontierSWE, at roughly a sixth of GPT-5.5's price.
Puppeteer any character, no rig
SCAIL-2, also from Z.ai, is an MIT-licensed model that animates a character straight from a driving video — no skeleton or pose rig — and handles identity swaps, multi-character scenes, even animal drivers at up to 704p. Weights are on Hugging Face, and a community Apple-Silicon (MLX) port is already in progress, so a Mac-based puppeteering tool is a real weekend build.
Copilot's CLI agents go multi-repo
GitHub Copilot CLI v1.0.73 (Jul 21) makes custom agents work reliably across multiple directories and keeps their instructions from drifting — useful for monorepos and multi-service checkouts. It lands right after v1.0.72's fix that stopped command approvals leaking between repos.