GLM-5.3-Flash: open MIT multimodal weights, 1M context, flash price
Z.ai's first natively multimodal GLM self-hosts under MIT, ~10x cheaper than 5.3; the 50% discount ends Sep 9. Plus Vercel budgets and gh CLI uploads.
Copy markdown
One model, images and video native
GLM-5.3-Flash is a 320B-parameter mixture-of-experts with only 18B active, and the first GLM-5 that takes text, image, video, and files without a bolted-on vision adapter. It reads up to 1M tokens of context.
MIT weights you can self-host
The weights sit on Hugging Face at zai-org/GLM-5.3-Flash under a plain MIT license — commercial use, no strings. With just 18B active params it's far lighter to serve than the 753B text-only GLM-5.3.
Flash pricing, on a discount clock
List API pricing is $0.15 / $0.50 per 1M input/output ($0.03 cached) — roughly 10x cheaper input than text GLM-5.3 at $1.40 / $4.40. A 50% launch cut halves it to $0.075 / $0.25 through Sep 9, 16:00 UTC.
The benchmarks it claims — and the caveats
Self-reported gains over GLM-5.2: DeepSWE 46.2 to 63.4, AutomationBench 26.2 to 48.8, plus vision scores of 78% Chartography and 62.4% OfficeQA-Pro. These are vendor numbers — verify before you commit. Max output tokens also differ by host (OpenRouter 131k vs HF 163k).
Elsewhere: Vercel AI Gateway adds per-user budgets
AI Gateway budgets now scope to individual team members, not just team, project, or key, so you can cap one engineer's model spend. Same drop: the AI SDK's xAI provider handles tool calls in batch, and Next.js 16.3.4 re-enables AVIF image optimization.
Elsewhere: gh CLI 2.99 uploads media, adds worktrees
GitHub CLI 2.99.0 adds an --attach flag to push images and video (up to 10MB images, 10-100MB video) straight into issues, PRs, and comments, plus git-worktree support in gh issue develop --checkout --worktree.