Goldman Sachs initiated coverage on Zhipu, the Hong Kong-listed parent of Z.ai also trading as Knowledge Atlas Technology, at a HK$1,880 price target on July 10, roughly 15% above the close. The two-report package the same day named Z.ai's GLM-5.2 alongside DeepSeek and ByteDance as the desk's preferred Chinese AI stack. That's a notable sequencing: a sell-side buy thesis built around an MIT-licensed, open-weights model whose HuggingFace card went live on June 13.

The benchmarks are what earned the coverage. On Terminal-Bench 2.1, GLM-5.2 posts 81.0, trailing Anthropic's Opus 4.8 at 85.0 and GPT-5.5 at 84.0 but clearing Google's Gemini 3.1 Pro at 74.0. On Artificial Analysis's AA-Briefcase leaderboard, Anthropic Fable 5 takes first at 1,587 Elo and Opus 4.8 second at 1,356, with GLM-5.2 third at 1,266. On Design Arena, Z.ai sits at the top at 1,360.

The architecture is where the model earns its unit economics. GLM-5.2 is a 744-billion-parameter mixture-of-experts checkpoint activating around 40 billion parameters per forward pass, with a 1-million-token context window, five times GLM-5.1's 200K, and a 131,072-token maximum output. An IndexShare sparse-attention scheme, one indexer shared across every four layers, cuts per-token compute at full context by roughly 2.9x. The full BF16 checkpoint is 1.51 TB.

The pricing is the part that matters for the Goldman thesis. Z.ai's API lists $1.40 per million input tokens and $4.40 per million output, versus GPT-5.5's $5 and $30. On output-heavy workloads that's about one-sixth the cost. Cached input runs $0.26 per million. A separate GLM Coding Plan tiers at $12.60, $50.40, and $112.00 per month billed annually.

Yuchen, a Databricks MTS, flagged the release on X, and Lambda circulated it through alphaXiv. The cultural read matters: an MIT licence removes the usage-restriction argument U.S. buyers have used to justify avoiding Chinese weights, and the Tom's Hardware benchmarking on Huawei silicon lands against the backdrop of ongoing U.S. Commerce Department export controls.

Open-weights releases have flirted with frontier scores before. The DeepSeek moment in early 2025 rattled markets on cost narrative alone. What's new here's the sell-side infrastructure catching up: Goldman is telling institutional clients to own the equity behind an MIT-licensed model that undercuts OpenAI's list pricing by a factor of six. The re-rating of the open-weights thesis stopped being a research-desk curiosity the moment a HK$1,880 target got printed next to it.

Goldman initiates on Z.ai at HK$1,880; GLM-5.2 scores 81.0 on Terminal-Bench 2.1

Sources