Maximilian Schreiner is Managing Editor at THE DECODER and has covered artificial intelligence for close to a decade, from new models and tools to their impact on businesses, media, and society. He brings that expertise to heise KI PRO, where he makes AI accessible for professionals in the DACH region - clear, critical, and practical. With a background in philosophy, one question keeps driving him: whether machines can truly think, or are just getting better at faking it.
Meta has released Muse Glimmer, the first open model from its new Superintelligence Labs. It’s a 30B agent model that runs on consumer hardware once the weights are compressed, needing less than 20 GB of memory. In an accompanying essay, Mark Zuckerberg mounts an aggressive defense of distilling other labs’ models and calls for fewer restrictions on US labs, a direct counterpunch at OpenAI and Anthropic. An open-weight version of Muse Spark 1.2 should follow soon, according to the Wall Street Journal.
Bytedance is training an AI model with up to ten trillion parameters, according to the Financial Times. That's three times the size of Moonshot's Kimi K3, currently the largest Chinese model. It would put the TikTok parent company in the same ballpark as Anthropic's top system Mythos 5, which industry estimates place at around eight trillion parameters. Anthropic hasn't disclosed its own numbers.
Three insiders told the FT that the model is in pretraining, a phase that typically takes three to six months. Parameters determine how much a model can store, but performance also depends on data quality and training methods. One of the sources says Bytedance has avoided distillation, meaning training on outputs from other companies' models, for over a year.
Founder Zhang Yiming told the 2,000-person Seed team internally to aim for world-leading model capabilities over the long term. xAI is also training Grok variants with six and ten trillion parameters on its Colossus 2 cluster, according to Elon Musk.
Alibaba's Qwen3.8 Max scores 56 on the Artificial Analysis Intelligence Index, a 10-point jump over Qwen3.7 Max (46). According to Artificial Analysis, that puts it on par with Claude Opus 4.8 and ahead of GLM-5.2 (51), but behind Kimi K3 (57), which also runs 25 percent cheaper.
On GDPval-AA, a benchmark for work-related tasks, Qwen jumps 468 Elo points to 1,739, passing Kimi K3 (1,685). Only Claude Opus 5 (1,852) scores higher. The catch is how it gets there. Qwen3.8 Max needs 64 steps per task instead of 14, and input tokens grew 15x because the test resends the full conversation history to the model at each step.
Image: AA
The model works more thoroughly but runs slower and costs more. Alibaba's price-to-performance ratio takes a hit despite lower token prices (input dropped from $2.50 to $2.00 per million tokens, output from $7.50 to $6.00, and cache hits from $0.50 to $0.25). A single task in the Intelligence Index now costs $1.14, more than double Qwen3.7 Max ($0.53). Kimi K3 scores one point higher at just $0.86 per task, and GLM-5.2 comes in at $0.57.
There are also regressions compared to the previous version. AA-LCR dropped 2 points, a test that checks whether a model can correctly pull together information from very long texts. AA-Omniscience fell 10 points, measuring whether a model answers knowledge questions correctly or honestly admits it doesn't know. The accuracy rate stays around 31 percent, but the hallucination rate jumped from 23 to 40 percent. Qwen3.8 Max guesses far more often instead of saying it doesn't know.
SpaceX plans to more than 5x its compute capacity by the end of 2027, betting exclusively on Nvidia’s Vera Rubin platform. The expansion could require well over a million new GPUs. Meanwhile, the company’s AI segment posted $2.56 billion in Q2 revenue, driven mostly by leasing out its own server capacity.
A US appeals court has overturned Amazon’s injunction against Perplexity’s AI shopping agents, ruling that it’s the users who access Amazon, not the startup. It’s the first federal appeals court decision on whether AI agents can lawfully act on online platforms on behalf of users, and it could reshape the entire AI agent industry.
The Trump administration discussed sanctions and cloud bans targeting Chinese open-weight AI models, according to the New York Times. OpenAI and Anthropic pushed for restrictions, while Nvidia, Google, and Meta fought back. After pushback from Silicon Valley, Washington backed off for now, but a decision is expected before Xi Jinping’s visit in September.
MiniMax releases H3 video model weights, putting an open model at the top of a video ranking for the first time.Artificial Analysis ranks H3 first in Video Editing, second in Text-to-Video, and third in Image-to-Video. The 33-billion-parameter model processes text, images, video, and audio together, generating four- to 15-second clips with stereo sound. According to the model card, a single prompt can include up to nine reference images, three video clips, and three audio clips.
Video by MiniMax H3
Two pieces remain closed, though. The 2K resolution module and H3-Context-IR, which translates prompts and reference material into a structured intermediate format, aren't included. Running H3 locally in ComfyUI tops out at 768p, and users will need to handle context prep themselves using MiniMax's published prompting guides. The open weights do allow fine-tuning on custom footage, characters, or a specific visual style. One catch on the license side: commercial use is only permitted for companies making under $20 million in revenue.
ByteDance released its closed Seedance 2.5 the same day, which generates 30-second clips with built-in audio.
