Every published Top AI Stories item tagged with Zhipu AI, newest first.
The September update of Mozilla's State of Open Source AI report puts the best open-weight models about 4.4 months behind the closed frontier on METR's task-horizon data, and roughly three points behind the closed leader on the Artificial Analysis Intelligence Index at a fraction of the price. On a neutral harness from Vals AI, Zhipu's GLM 5.2 landed within a point of Claude Opus 4.7 and 4.8 on Terminal-Bench 2.1 at about a fifth of the cost per completed task. Mozilla's rule is workload-specific: closed models earn their premium on tasks that take a human expert eight to twelve hours, on expert professional work and on long context, while routine work belongs on open models. Eight of the ten highest-volume models on OpenRouter in August shipped open weights, and nearly all of the leading ones are Chinese, which Mozilla's Raffi Krikorian calls "a plurality and a concentration at the same time."
Z.AI, the Beijing lab formerly known as Zhipu AI, raised $5 billion through a Hong Kong share placement worth roughly $2 billion alongside a convertible bond sale of about $3 billion, according to an exchange filing made Sunday. The new shares priced at a 10 percent discount to Friday's close. Roughly 60 percent of the net proceeds are earmarked for research on next-generation models and self-training systems, with a further 15 percent for expansion; the company listed in Hong Kong in January and raised about $4 billion in a follow-on sale in July.
A new Anthropic threat intelligence report attributes distillation campaigns with high confidence to Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax. A joint advisory from three US agencies asserted this pattern two days earlier; what is new here is the platform's own telemetry. Alibaba ran more than 151 million exchanges between May and July, peaking near three million a day across thousands of fraudulent accounts, to capture the chain-of-thought reasoning behind Opus. Moonshot and DeepSeek went further, silently routing their own customers' requests to Claude and keeping the transcripts, exposing names, live credentials and corporate data to a third party those users never chose.
The National Security Agency, the Federal Bureau of Investigation and the Cybersecurity and Infrastructure Security Agency issued a joint advisory saying DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai have extracted billions of tokens from Claude, GPT, Gemini and Grok since late 2024, likely with Chinese government awareness. That escalates a July accusation from the White House technology chief, which named Moonshot alone. Distillation, training a smaller model on a larger one's outputs, is a standard technique — but the agencies say it forms the core rather than a supplement of these labs' development. They also reject the number behind the 2025 market shock, DeepSeek's widely cited training cost of 5.6 million dollars. The advisory says it leaves out the cost of the data taken this way. China's foreign ministry calls the claims groundless.
Two weeks after announcing the model, Z.ai has published the GLM-5.3 flagship on Hugging Face — ungated, 153 files, roughly 753 billion parameters. The company had delayed the drop for safety evaluation after, in its own words, cyber capability developed faster than expected as post-training scaled. Its benchmark table puts GLM-5.3 first on CyberGym vulnerability discovery at 84.5 percent, ahead of Claude Fable 5 and GPT-5.6 Sol, though it still trails both by a wide margin further up the exploitation chain. The license is not the plain MIT that GLM-5.3-Flash carried this week: it adds a security review by Z.ai for any model-as-a-service operator whose revenue passes 10 billion dollars over twelve consecutive months.
Z.ai told Bloomberg that Ox Alpha — the unnamed, free, frontier-class model that had been serving traffic on OpenRouter while nobody would say who made it — is the newest model in its GLM line, and said weights would follow. They have: GLM-5.3-Flash is now on Hugging Face, ungated, 63 weight files, under a plain MIT license with no revenue threshold and no regional carve-out. That is a notably permissive answer from a lab that two weeks ago held a release back on the grounds that its cyber capability had grown faster than expected — though the flagship GLM-5.3, whose weights were the ones actually delayed, still has no public repository.
A model called Ox Alpha went live on OpenRouter on August 20 as a stealth listing, free to use, with a context window of about 1.05 million tokens and text, image and video input. OpenRouter's own page says only that it is "developed and operated by a third-party provider who has chosen to remain anonymous during this preview," and that prompts and completions are retained by that provider, though not used for training. Developers are therefore sending working code to a party they cannot name. Community fingerprinting points at Z.ai's GLM family; Zhipu has not commented. Stripe chief executive Patrick Collison, whose company is buying OpenRouter, called the model "very impressive."
Z.ai released GLM-5.3 through its API on August 14, keeping the same base model as GLM-5.2 and taking every gain from scaled-up post-training. The company reports 84.5 percent on CyberGym, edging past Claude Mythos 5 and GPT-5.6 Sol, and says the model began reasoning across multiple stages of exploitation rather than isolated bug-finding. Z.ai is holding the open weights back for about two weeks of safety evaluation before anyone can download them.
Microsoft chief executive Satya Nadella argued that companies buying proprietary AI "pay twice" — once in fees, and again in the proprietary know-how they hand over through prompts, tool calls, and corrections that quietly train someone else's model. His warning lands as the Financial Times reports a wave of firms — Coinbase, Shopify, and Airbnb among them — shifting to open-weight Chinese models like DeepSeek, Alibaba's Qwen, and Zhipu's GLM, which can run 60 to 90 percent cheaper. Coinbase says it nearly halved its AI spending even as usage climbed.
China's Ministry of Commerce has held meetings with Alibaba, ByteDance, and startup Zhipu AI about whether to limit foreign access to the country's most capable models, according to Reuters — a striking reversal for labs whose open-weight releases have been the main challenge to US frontier dominance. Options sketched in the talks reportedly range from security reviews to barring the most sensitive models from public release, and cover open-weight systems like Qwen, Doubao, and GLM, not just proprietary ones. Nothing is decided, and officials have made no public comment.
Analyst Martin Alderson argues Zhipu's open-weights GLM 5.2 is the first open model good enough to stand in for Claude Opus on many tasks — at roughly a fifth of the price, near four dollars and forty cents per million tokens, and nearly 30 times cheaper than GPT-5.5. His thesis: frontier labs run about 90 percent gross margins on inference, and with switching costs this low, drop-in open models threaten that entire business model. It is a sharper version of the compute-economics worry that has surfaced repeatedly this quarter — this time driven by open weights rather than power bills.
Zhipu AI released the full MIT-licensed weights for GLM-5.2, the coding-first model it teased earlier this month, and independent benchmarks now back the hype. The 753-billion-parameter mixture-of-experts model (about 40 billion active) scored 62.1 on SWE-bench Pro — beating GPT-5.5 — and finished within about a point of Claude Opus 4.8 on several long-horizon coding tests, at roughly one-sixth the price. It now tops the open-weights intelligence rankings, a fresh sign that China's open models are closing the gap with frontier US labs.
China's Zhipu AI released GLM 5.2, a coding-first model with a 1 million token context window, pushing it live across all four tiers of its GLM Coding Plan. The company is positioning it as a permissively licensed alternative to Claude Code and GPT-5.5 for the Asia-Pacific market, with a standalone API and MIT-licensed open weights promised for next week. One caveat: Zhipu published no benchmarks at launch — no SWE-bench, no LiveCodeBench — so its "powerful coding" and "long-horizon" claims stay unverified for now, echoing the skepticism that met Moonshot's Kimi release a day earlier.