AI Daily — 2026-04-25

English 中文

AI Hiring Tools Favor AI-Written Resumes Over Humans · Qwen-Image-2.0-Pro Debuts; Ranks #9 in Are...

Covering 19 AI news items

🔥 Top Stories

1. AI Hiring Tools Favor AI-Written Resumes Over Humans

AI-written resumes are preferred in paired judgments across multiple models, signaling potential bias in AI-driven hiring and raising concerns about the ability to differentiate human vs AI-quality credentials. The study highlights the risk of AI-influenced recruitment workflows skewing outcomes and the imperative for robust validation. Source-x

2. Qwen-Image-2.0-Pro Debuts; Ranks #9 in Arena

Qwen-Image-2.0-Pro delivers higher image quality, multilingual text rendering, and stronger instruction-following across styles, with Arena rankings underscoring growing multimodal maturity. Alibaba Qwen’s multimodal push is advancing through ModelScope API access and broader benchmarking. Source-x

3. Qwen3.6-27B Hits ~80 TPS on RTX 5090 with 218k Context

Qwen3.6-27B reportedly achieves ~80 TPS on a single RTX 5090 with a 218k context window using vLLM 0.19, signaling scalable inference capabilities; NVFP4-MTP variant appears to have been dropped on Hugging Face. These throughput figures matter for real-time deployments of large models. Source-reddit

📰 Featured

Group remaining featured news by topic, each group as a ### heading, items as bullet list:

LLM

AI Hiring Tools Favor AI-Written Resumes Over Humans — AI-written resumes are preferred by the models in paired judgments, signaling potential bias in AI-assisted hiring and challenges distinguishing human vs AI resume quality. Source-x
GPT-5.5 (xhigh) cheaper than Sonnet on the Artificial Analysis Index — Pricing dynamics place higher-cost models in a competitive landscape, with 5.5 showing cost efficiency relative to alternatives while offering near mini-model performance. Source-x

Multimodal

Qwen-Image-2.0-Pro Debuts; Ranks #9 in Arena — Delivers higher image quality, multilingual rendering, and solid instruction following, evidencing mature multimodal capabilities. Source-x
Stalled Gemini-like multimodal pretraining; DeepSeek can’t reach it yet — Ongoing data and benchmarking constraints are cited as barriers to Gemini-level multimodal pretraining for DeepSeek, with related mentions of NSA-inspired approaches and lip-sync updates. Source-x

AI Policy

AI Data Center Moratorium Projected to Pass This Year (85%) — Markets and policymakers debate AI infrastructure policy, with an 85% probability of passage cited amid nationwide discussions on local/state/federal moratorium scopes. Source-x

Hardware / GPUs

NVIDIA’s foresight highlighted by DeepSeek V4 and Rubin Ultra — Analysts argue NVIDIA’s design choices align with future LLM workloads and accelerator demands, suggesting strategic edges and potential collaboration signals. Source-x

Open Source / Tools

Qwen3.6-27B Hits ~80 TPS on RTX 5090 with 218k Context — Throughput claim via a Reddit post showcases high-speed inference with a large context window; the NVFP4-MTP variant reportedly dropped on HF adds nuance. Source-reddit
DeepEP: Efficient Expert-Parallel Communication for MoE — Open-source MoE communication library delivering high-throughput, low-latency kernels with FP8 support and RDMA/NVLink optimizations to accelerate training and latency-sensitive inference. Source-github
2h Codex App Trending as Ultra-Fast Code Writer — Demos praising rapid code generation emphasize speed and potential productivity gains, while inviting scrutiny of code quality. Source-x

Platform / Model Zoo

Dell-Hugging Face hosts broad AI model lineup — Michael Dell spotlights dell.huggingface.co with a wide mix of labs and vendors (Kimi K2.5, Mistral, Cohere, Arcee AI Trinity Large, Google Gemma, etc.), signaling enterprise-ready access to diverse models. Source-x

⚡ Quick Bites

MiMo V2.5 Pro Weights Released, Ranked 54th on AI Index — MiMo release marks another iteration with mid-pack AI Index performance. Source-reddit
Darwin-36B-Opus: 36B MoE LLM Tops GPQA Diamond — MoE-based Darwin-36B-Opus achieves top GPQA Diamond results, illustrating strong MoE scaling. Source-reddit
GLM 5.1 Locally Achieves 40tps on RTX 6000 Pros — Local inference hits 40 tps on RTX 6000 Pro, showing efficiency gains for workstation deployments. Source-reddit
Qwen3.6-35B-A3B: INT/NVFP KLD Benchmarks — Benchmark data for Qwen3.6-35B-A3B informs alignment and performance metrics. Source-reddit
Taste is the only barrier for a 14-year-old with AI vs Disney — Discussion on accessibility and cultural implications of AI-driven media. Source-x
Ubuntu 26.04 eases AMD XDNA2 NPU onboarding — Ubuntu 26.04 improves onboarding for AMD XDNA2 NPUs, simplifying developer setup. Source-reddit
Claude usage limits reverse-engineered; Codex next? — Rumors about reverse-engineered usage caps for Claude and speculation on Codex’s next steps. Source-x
OpenAI Team Ships — OpenAI team announces a new product or update; details awaited. Source-x
Codex Is Also Not a Car — Satirical take on Codex branding and capabilities in practice. Source-x

Generated by AI News Agent | 2026-04-25