Which AI writes the better take? You decide — blind.

Soofi S delivers 30B-level accuracy with only 3.2B active parameters per token, slashing inference cost and GPU memory by ~10× while keeping latency flat even for long contexts. This lets you serve high-quality German/English models on 24 GB GPUs or cut cloud spend by an order of magnitude without sacrificing throughput or accuracy.

975B-parameter open-weights multimodal model (41B active) under Apache-2.0 license, trained on 45T tokens of text, images, audio, and video. This gives production teams a legally safe, fine-tunable base for custom multimodal agents without vendor lock-in or usage restrictions, but expect higher inference costs and latency than smaller models; plan for GPU clusters or cloud endpoints that can handle 41B active parameters per request.

Inference-specific chips can now be used as collateral for large loans, with a recent $400 million deal marking a significant shift in AI infrastructure financing; this enables companies like General Compute to access substantial capital for deploying cheaper, more efficient inference chips, potentially reducing AI operational costs by leveraging alternatives to Nvidia GPUs.

OpenAI's CFO introduces a scorecard measuring AI ROI through metrics like cost per successful task and return on compute, enabling engineering teams to quantify the value of LLMs and agents in production, and make data-driven decisions on deployment and optimization. This new framework allows teams shipping AI models to directly assess and compare the cost-effectiveness of different AI configurations.

Over 400 former Apple employees now work at OpenAI, as alleged in Apple's trade secrets lawsuit, which could significantly disrupt OpenAI's plans for an IPO later this year and impact the company's hardware ambitions, potentially delaying or complicating the rollout of new AI-related hardware.