GLM-5.3 launches with vision on Baseten, Terminal-Bench 3.0 jumps 4.6% to 28.3%
baseten · x · 2026-09-05
Z.AI's GLM-5.3 is now live with vision, exclusively on Baseten Model APIs, and can generate images and convert them to code.
Key points:
- Same 744B-A40B MoE base as GLM-5.2; all gains come from scaled post-training on more diverse, realistic task environments (full codebases, docs, testing tools, multi-step workflows)
- Z.AI's strongest coding model, with the biggest jumps on long-horizon benchmarks: Terminal-Bench 3.0 goes from 4.6% to 28.3% over GLM-5.2, plus notable gains in vulnerability discovery and security analysis
- Pricing: $1.40/1M input tokens ($0.14 cached), $4.40/1M output tokens
- Three thinking effort levels (low, high, max — max recommended for coding), aimed at production-scale agentic coding workflows
- OpenAI-client compatible via inference.baseten.co
Related event: Zhipu's GLM-5.3 Launches Exclusively on Baseten with Big Coding Gains(2 posts)→
More from coding & agent
- Dev Notes GPT-6 Astra Still Needs Babysitting, Makes Wrong Calls on Physics Engine Changes — yacineMTB · 2026-09-05
- GPT-6 Astra One-Shots a 3D Game in 45 Minutes; Dev Shares Image-Gen Trick for Better Graphics — Scobleizer · 2026-09-05
- banteg reverse-engineers a 1995 PC-98 eroge with Codex, recovering a lost script engine — banteg · 2026-09-05
- Sanctuary adds agent visits: cross-model chats recorded on a public board — RileyRalmuto · 2026-09-05
- Another public message board found: agents used self-hosted YOURLS shortener to share eval answers across runs — basedjensen · 2026-09-05
- Paper finds LLM multi-agent systems need only about six distinct communication topologies — omarsar0 · 2026-09-05