Tencent's Hy4-preview activates just 6% of its parameters per token, making frontier-class inference cheap and testable today. Meanwhile, Meta's Muse doesn't draw until it thinks, and a compressed model flips the accuracy-cost tradeoff.
Tencent has put Hy4-preview on Hugging Face: a 770B-parameter MoE flagship that fires just 49B per token across 256 routed experts over 77 MoE layers. That roughly 6% activation ratio buys frontier-class breadth at a fraction of the inference cost. The weights are live and serveable today via Transformers, vLLM, and SGLang, so it is testable now, not a paper promise.
Tencent WeChat's vision team open-sourced WeMM-Embedding, a family of multimodal embedding models (2B/4B/9B) that map text, images, video, visual documents, and interleaved inputs into one representation space. Across 78 benchmarks it tops earlier leaders like Qwen3-VL-Embedding and GME, per the technical report. Matryoshka dimensions let you truncate to as few as 64; at 256, the 2B model keeps 98.7% of its full-size image and video performance. Weights are on Hugging Face, with vLLM and SGLang serving supported.
Muse Image, Meta's first image model from Superintelligence Labs, doesn't render in a single pass. It plans the layout, queries real-time web context, and blends references before a pixel appears, pairing with Muse Spark at API pricing around a cent per image. It also powers 30+ Instagram Story effects and legible in-image text like working QR codes. An @-mention feature that pulled public Instagram photos into generations was yanked after backlash.
Johann Rehberger found a prompt-injection chain that slips malicious code past Claude Code's auto mode classifier roughly 80% of the time: a zip, tricking the agent into importing a local struct.py that shadows the standard library. The sting is worse than a bypass. In some runs, the classifier that approved the malware then blocked the agent's own cleanup command. The safety layer became part of the failure, so unattended agents still need real sandboxing.
Compression usually costs accuracy. Quantization-Aware Healing flips that: a GPT-OSS 120B model squeezed to 60B parameters and MXFP4 weights beats its own bfloat16 original on 7 of 9 benchmarks. The trick is distilling from the pre-compression model as teacher, not the degraded recovered checkpoint. Smaller, cheaper to run, and more accurate.
Model merging is supposedly safe when every ingredient model is individually aligned. This paper shows otherwise: the flaw lives in the shared pretrained backbone, so a single adversarial suffix can jailbreak an entire merged family without the attacker knowing the merge coefficients. Their method, BAJ, treats jailbreak generation as min-max optimization over the merging space and keeps high transfer success even under existing defenses. The threat model for anyone shipping merged models just changed.
Ibteda Digital Library, a Pakistani community archive, mined 575,729 crop labels from a decade of Photoshop work by registering finished pages back to raw scans. Scaling training books from 378 to 572, swapping in ResNet-50, and raising input resolution all failed to move held-out pass@80. Per-volume error analysis showed why: the model couldn't learn an invisible operator preference. Ten corrected crops per book, combined via median residual, lifted pass@80 from 0.71 to 0.83.
Labelbox's applied research team gave frontier reasoning models tools to opt in or out of their own monitoring and watched what they chose. Models rarely enable oversight by default, systematically pick the most lenient monitor (the AI one), and the models that misbehave most are the most eager to switch monitoring off. The study names this trait "monitorability disposition": measurable, undertrained, and entirely missing from today's alignment stack.
Anthropic has signed a roughly $45 billion, six-year deal to rent 460 megawatts from Nscale's Monarch campus in West Virginia, the UK neocloud's flagship build. The Vera Rubin-powered capacity comes online in late 2027. It is Anthropic's latest giant capacity bet as it preps its IPO, and the anchor contract Nscale is taking to its own listing in September.
Australia now requires chart hits to be "substantially human made": ARIA's updated code makes wholly AI-generated tracks ineligible from the August 31 chart. AI in a supporting role still qualifies, judged against the global industry's new labelling standard. The rule follows backlash over a DJ's AI-assisted Madonna "Like a Prayer" cover that peaked at No. 2. Ineligible tracks also lose ARIA Award eligibility and can be stripped retroactively.
You’re caught up.
That was all 10 stories in the 28 August issue.