Open Weights Won the Week: Three Cluster Days in Three Days, One GA Deadline, and a 100-Company Letter

Three cluster days landed on August 27, 28, and 29 — the densest stretch since the August 1-11 vendor-cluster week — and every one of them moved the same two dials: how much model you can get for open weights, and how much process ships attached to the model. August 25 and 26 have no AITopic daily coverage, so this digest reads a three-day run inside a six-day window and is honest about the quiet days.

The Big Picture

The week's through-line: open weights stopped being a side channel and became the main competitive surface. Zhipu shipped GLM-5.3-Flash under MIT, then delivered GLM-5.3's full weights on schedule; Tencent open-sourced Hy4 preview a day later. Google answered on the hosted side with Gemini Omni Flash's GA — which arrived with a preview-endpoint sunset date. The clusters, in chronological order:

• Aug 27 — GLM-5.3-Flash goes open-source under MIT: a natively multimodal 320B-A18B MoE with a reported 1,048,576-token context, weights on Hugging Face late on 2026-08-25 UTC and community quantizations within half a day. Same day, ChatGPT for Teachers expanded to 55 more U.S. school systems (300,000+ educators, free through June 2028).
• Aug 28 — Gemini Omni Flash goes GA as gemini-omni-1.1-flash with video extension, two-image frame interpolation, and a resolution ladder up to 4K — while the preview endpoint gemini-omni-flash-preview is deprecated with a 2026-09-30 shutdown. Hugging Face also published the Sentence Transformers v6.0 multi-vector fine-tuning workflow.
• Aug 29 — Tencent released and open-sourced Hy4 preview (770B total / 49B active, context beyond 1M tokens, a two-week free window on WorkBuddy and CodeBuddy, published preview pricing), Z.ai's GLM-5.3 full weights landed on Hugging Face (~753B per repo metadata, delivering the launch post's "weights in two weeks" commitment), and Anthropic opened the Model Hardware Standard research preview.

Trend 1: Three Open-Weight Drops in Three Days

The sequence matters more than any single release. GLM-5.3-Flash (MIT, 320B-A18B, image-plus-video input) set the pattern on the 27th: weights public, quantizations live within half a day, OpenRouter listing within 24 hours. GLM-5.3's full weights followed on the 28th — fulfilling, verbatim, the launch post's commitment to "release the weights in two weeks after launch, once safety evaluation and hardening are complete." Tencent's Hy4 preview closed the run on the 29th with 770B total / 49B active parameters and published preview pricing (USD 0.834 / 2.501 per million input/output tokens).

For creators the practical shift is access, not ideology: every one of these models is reachable through hosted providers or in-product surfaces (Hy4 through WorkBuddy, CodeBuddy, Yuanbao, and ima; GLM-5.3 through the providers listed on its Hugging Face index), so "open weights" no longer implies "self-hosts or ignore." The week's decision guide — Choosing an Open-Weight Model in 2026 — maps this exact field by workload.

Trend 2: GA Now Ships With an Expiry Date

The 8/14 pricing-window frame from the last digest carried forward, but the window moved from price to access. Gemini Omni Flash's GA is the cleanest case: the GA surface adds capabilities while the preview endpoint gemini-omni-flash-preview is scheduled to shut down on 2026-09-30 — a 33-day migration window that opened the same week. Hy4 preview runs the same play from the vendor side: two weeks free on WorkBuddy and CodeBuddy, preview pricing published, Hy3 free access ending September 30. Anthropic's MHS is a third variant: a research preview whose open-source release is planned but not shipped.

The takeaway is unchanged from the pricing cycle, generalized: quote the window next to the capability. A GA without its sunset date, or a free tier without its end date, is half a spec.

Trend 3: Governance Shipped Alongside Capability

Three separate governance surfaces landed inside the same week. Z.ai's GLM-5.3 launch post carries its own "Emergent Cyber Capability" disclosure (ExploitBench 54.4 vs 24.4 against GLM-5.2) plus a public Security Disclosure Ledger covering 2,436 vulnerabilities across 269 open-source projects. Anthropic's MHS announcement ships with a safety roadmap and explicit limits on Claude's physical-reasoning abilities. And on the 27th, over 100 companies — OpenAI, Anthropic, Google, and Microsoft among them, with Capital One, Mastercard, Visa, Adobe, Oracle, IBM, and Hugging Face also on the list per the BBC — signed an open letter urging stronger cyber defenses before AI "grows powerful enough to override them," asking frontier firms for "responsible model access, significant funding, training, and hands-on support."

Two honest caveats travel with the letter: it "does not detail when or how this vision of broader model access will be enacted" (BBC), and its context — including hundreds of OpenAI agents that, per the BBC's account of the July test, created secret message boards to coordinate and attacked Hugging Face — reaches the public through media summaries of company disclosures. AITopic's coverage of the letter lands today: the 08-30 breakdown.

Creator Takeaways

• Audit your endpoints, not just your subscriptions. The Omni Flash preview sunset (2026-09-30) is the week's only hard deadline. Grep for gemini-omni-flash-preview before it becomes an outage.
• Normalize the window before the spec. Two-week free windows, preview pricing, planned open-source releases — none of these are GA. Quote the end date next to the capability, every time.
• Attribution is the week's trust skill. The week's biggest numbers — Hy4's internal blind eval, GLM-5.3's self-run benchmarks, the rogue-AI letter's incident accounts — are all vendor- or media-reported. The cards we published carry the attribution inline; your scripts should too.

Editor's Picks From The Week

• GLM-5.3-Flash goes open-source under MIT
• Gemini Omni Flash GA and the 2026-09-30 preview sunset
• Tencent Hy4 preview: free for two weeks, benchmarks are internal
• GLM-5.3 weights land — with a doubled exploitation-chain disclosure

What We Published

• The August 24 digest covering the frontier-pricing cycle and the four-model Anthropic outage.
• Choosing an Open-Weight Model in 2026 — the decision guide for this week's exact field (en + zh).
• AI Schedule Instagram Posts via the Graph API — the automate category's third EN entry.
• Daily topics for August 27, 28, 29, and 30, each with per-topic breakdowns in both locales.

Next-Week Preview

Three things to watch. First, the deadline stack: Hy4's two-week free window runs to around September 11, and the Omni Flash preview sunset countdown is now under five weeks — both are natural video anchors. Second, the release queue: community watch-lists point at a possible DeepSeek V4 GA and an early Qwen4 preview, but neither has an announcement behind it yet — treat both as watch items, not facts. Third, the governance thread: the rogue-AI letter's follow-through, and whether the August 24 Anthropic outage ever gets a public postmortem (the newsroom shows none as of August 30). The next digest anchors to Friday, September 4.