curated · 2026-06-01

Zhipu’s STAR Market plan is a funding handoff, not proof that its model business has hardened. The filing says new A-shares will be 2% to 8% of post-issuance equity, with proceeds for foundation models, a MaaS platform, and working capital. IT Home’s linked coverage lists 2025 revenue at RMB 724 million and adjusted net loss at RMB 3.182 billion. That ratio is the whole tension. I don’t buy the clean “commercialization leader” framing here. Zhipu has GLM, AutoClaw, and government-enterprise MaaS channels, but public-market buyers inherit compute spend, slow enterprise sales, and margin pressure from DeepSeek-style open-source pricing anchors. The rename to Z.AI smells like capital-market packaging as much as product clarity.

MiniMax dropped M3, an open-weight model that does three things at once: 1M token context window, native multimodal input (images and video), and coding performance that claims to beat GPT-5.5 and Gemini 3.1 Pro on several benchmarks. Both sources are pulling from the same official blog post, so we're looking at one voice here, not independent confirmation. I'd take the benchmark numbers with a grain of salt for now. SWE-Bench Pro at 59% and Terminal-Bench at 66% are legitimately strong scores, but they're self-reported. The MSA attention architecture is the more interesting bit—at 1M context, per-token compute drops to 1/20 of their previous model, with 9x faster prefill and 15x faster decode. If those speedups hold in production, the cost story could be compelling. What's missing: API pricing, actual weight downloads, and any real-world usage reports. Open-weight models often look great on paper and turn into a headache to deploy. Give it a few days for the community to kick the tires before calling this a closed-source killer.