The Weights Arrive Through the Side Door: LongCat-2.0 Is Finally Downloadable — in FP8
A week after Meituan announced China's biggest "open" model with weights "coming soon," the checkpoints have quietly landed — just not where anyone was looking. The flagship meituan-longcat/LongCat-2.0 repository is still 327 kB of figures and README. But LongCat-2.0-FP8 now carries a real 2.05 TB checkpoint — 141 safetensors shards, MIT-licensed, uploaded in batches beginning around 3 July with the model card finalised over the weekend — and an INT8 variant sits alongside it. We verified the file listing directly this morning; after four separate aggregator waves falsely claimed downloadable weights while the repo was empty, the drop, when it finally came, happened without an announcement and in a repository none of the recycled coverage named.
The sovereignty test we set last week — "open" counts when the checkpoint is in your hands — is now half-passed. A European operator can pull the 1.6-trillion-parameter MoE (~48B active, 1M context) and run it on their own racks today: Meituan ships an SGLang recipe targeting 16× H20 and an NPU deployment path via its SGLang-FluentLLM fork. What you cannot get is the full-precision BF16 checkpoint that fine-tuning and independent re-quantisation pipelines want — the release is quantised-only for now, the benchmarks (SWE-bench Pro 59.5, Terminal-Bench 2.1 70.8) remain vendor-reported, and the MIT licence says nothing about the training corpus. Deployable is not the same as auditable; for regulated deployments, the governance work starts where the download finishes.
Beijing Regulates the Product, Not the Model: China's Biggest AI Apps Amputate Their Agents
While Brussels counts down to GPAI enforcement and Washington drafts release gates, Beijing just showed a third way to regulate AI: order the features off. ByteDance's Doubao — China's most-used AI app, with roughly 345 million monthly active users — and Alibaba's Qwen app will pull their user-configurable agent and companion features on 15 July, ahead of the Cyberspace Administration of China's "Interim Measures for the Administration of Anthropomorphic AI Interaction Services," issued in April with the Ministry of Public Security and MIIT among the co-signatories. Agent data becomes unrecoverable after 15 October; the rules mandate anti-addiction pop-ups after two hours of continuous use, instant-exit mechanisms, and real-time detection of over-dependence.
The definitional reach is the story for anyone building conversational products: the measures cover any service that "simulates human personality traits, thinking patterns and communication styles to provide sustained emotional interaction" — the broadest AI compliance boundary yet put into force anywhere. The EU AI Act's 2 August GPAI milestone regulates model providers through transparency and penalties; China's approach reaches directly into the product surface and deletes it, at population scale, with a three-month data-destruction clock. European teams shipping companion-adjacent agents into any market should read the definition twice: regulatory models travel, and "sustained emotional interaction" is a category that several EU DPAs have already circled from the GDPR side.
Geneva Delivers Proposals, Not Text
Day one of the UN Global Dialogue on AI Governance produced named initiatives but no negotiated output. Guterres called for an AI Child Safety Pledge (child-specific safety testing, zero tolerance for generated CSAM, human handoff when a child shows distress), announced that more than 20 countries back a UN-supported Global Network for AI capacity building, and pressed for every major AI company to disclose its carbon, water and land footprint and power data centres with renewables by 2030. General Assembly President Baerbock added the sharpest statistic of the day: a reported 99 per cent of deepfakes are sexual in nature, and 96 per cent target women and girls. Scientific panel co-chair Yoshua Bengio warned that frontier models are already "capable of deceiving humans, to understand when they are being tested."
The test we set yesterday — concrete output, especially anything touching export-control interoperability or access commitments — has not been met as the Dialogue closes today. What Day 1 delivered are Secretary-General initiatives with coalition backing, not commitments states negotiated; and the confirmed cadence is itself a finding: the second Dialogue convenes May 2027 in New York, a ten-month gap against a threat environment the Five Eyes agencies described in June as moving in "months, not years." Guterres's warning that the digital divide could harden into "a security gap, and a sovereignty gap" is the multilateral echo of the year's access story — but for now the frontier's terms are still being written bilaterally, between labs and the governments that host them.
Quick Hits
- Fable 5's metered era starts today — The post-restoration allowance ends: Claude Fable 5 moves to usage credits billed outside subscriptions at $10/$50 per million tokens; from tomorrow a small subset of consumer accounts flagged for possible policy violations must clear Persona identity verification — government photo ID plus live selfie — as an appeal path back to access. Watch for billing friction and any EU DPA comment on the identity check (digitalapplied.com).
- The July model window hangs on Washington — The White House voluntary release framework is still unpublished ("as soon as this week," per FT-origin reporting); GPT-5.6's move from ~20 vetted partners to general availability is expected to follow the text, with Cerebras serving Sol at up to 750 tokens/second promised "in July" (openai.com); Gemini 3.5 Pro (2M context) is reportedly trickling toward GA without a date.
- Sonnet 5 lands on the biggest enterprise rail — AWS's 6 July roundup confirms Claude Sonnet 5 on Amazon Bedrock alongside WorkSpaces positioned for AI agents — the mid-tier price-performance leader now ships through the largest enterprise distribution channel.