I've been using GLM-5.2 extensively over the past few days, primarily for software engineering tasks through OpenCode and Hermes-Agent, and it has genuinely impressed me.
The biggest compliment I can give is that, during many coding sessions, it gave me a similar feeling to using Claude Opus. Not because the models are identical, but because of how reliably it understands context, follows complex instructions, and completes multi-step engineering tasks.
I also tested it through Ollama Cloud using an API key, and the experience was consistently solid across different environments.
What I appreciate most is that the team decided to open-source such a capable model. The AI community benefits enormously when high-quality foundation models are openly available, and I wanted to leave a review simply to thank and recognize the work that went into this release.
Looking forward to seeing how the project evolves.
GLM-Legal is Z.ai’s legal AI model, built on GLM-5.3-Flash and trained for agentic legal work. That means multi-step tasks like contract review, evidence analysis and compliance research. It’s in an application-only API beta.
Problem: Legal work is long, detailed and unforgiving. A contract review or compliance check has many steps, and one missed point can sink the result. General-purpose models often fall short on that kind of multi-step legal work.
Solution: Z.ai took its GLM-5.3-Flash model and trained it for legal tasks. Per Z.ai, that meant legal mid-training on a larger corpus, then reinforcement learning across tens of thousands of legal task environments. It built the agent harness alongside the model, so the two were improved together.
How it performs:
Harvey’s Legal Agent Benchmark: 19.17% of tasks passed, up from 6.67% for the base model. A task only passes if every one of its criteria passes, so even the top scores are low in absolute terms.
LegalBench: 86.23%, versus 88.56% for Claude Fable 5. Z.ai says it matches frontier models here, but its numbers show it slightly behind.
Cost: about $0.15 per Harvey LAB test on Vals AI’s runs, much lower than the other models compared. The price Vals used isn’t a published GLM-Legal price.
Access:
API: It’s an application-only beta, with no published price
Weights: no public weights reported, and no Hugging Face repo yet
Desktop: Z.ai says it’s coming to its AutoClaw desktop agent app, with no date given
Who it’s for: law firms, in-house counsel and compliance teams exploring AI for legal workflows. This is a tool for professionals and doesn’t replace legal advice or a lawyer.
Read Z.ai’s announcement · See the Vals AI results · Apply for API access