GLM-5.3 is now open-weight: strongest open coding model to self-host
Z.ai's 744B MoE is downloadable and on 21 OpenRouter providers from $1.17/M — it leads open models on coding, and a cheap multimodal Flash ships too.
Copy markdown
The strongest open-weights coder, now downloadable
GLM-5.3 is a 744B mixture-of-experts (~40B active per token) that Z.ai calls its most capable open coding model — a ~50% jump over GLM-5.2 on its in-house Code Bench and 66.9 on DeepSWE v1.1. The weights are on Hugging Face, so you can fine-tune it or run it fully behind your own firewall.
It reportedly edges Opus 4.8 on SWE tasks
On FrontierSWE it posts 78.1 — reportedly ahead of Claude Opus 4.8 — plus 28.3 on Terminal-Bench 3.0 and 28.5 on Agents' Last Exam (CLI), the top open-weight marks on both. It still trails closed leaders like Fable 5 and GPT-5.6 Sol, so treat it as your best local coding agent, not a frontier swap.
Three ways to run it today
Self-host on 8xH200 at fp8 (~750GB), route to it across 21 OpenRouter providers at $1.17/$3.96 per million tokens in/out, or subscribe to the GLM Coding Plan from $12.60/mo. Context reaches 256K, with 1M on some endpoints.
The cyber scores are the real story — and the caveat
GLM-5.3 hits 84.5% on CyberGym for vulnerability discovery, which Z.ai says beats every closed model, and more than doubled GLM-5.2 on ExploitBench (54.4%). That makes it a strong defensive-security helper — and the reason Z.ai held the weights for a two-week safety review before opening them.
GLM-5.3-Flash: the cheap multimodal sibling
Alongside the flagship, Z.ai's native-multimodal Flash is open on Hugging Face and on OpenRouter at about $0.07/$0.24 per million with a 1.3M-token context — a near-free option for high-volume image-plus-text pipelines.
Elsewhere: Cohere's Parse 5 turns documents into Markdown
Cohere shipped parse-v5.0, a 2.3B vision-language model that converts enterprise PDFs and scans into clean Markdown, now live in Microsoft Foundry and tuned to win on cost per page. Useful if you're building RAG over messy document piles.