#benchmarks

Run LLMs on iPadOS with LemonSeed Studio’s new engine: 158 tps decode, 1.6k tps prefill using an AMD GPU. Unleash local AI on the go.

Manus secures $500M, fueling AI agents, new “Cue” app, and a potential $4B valuation—showing China’s AI startup resilience despite Meta fallout.

Micron’s NVHBM moves the memory controller to a base die, boosting bandwidth by 30%, cutting power 15% and shrinking I/O area 67%—yet high cost keeps it in enterprise‑AI markets.

AI now outperforms researchers in “experimental research taste” benchmarks, revealing both strengths and limits of automated scientific evaluation.

Google DeepMind’s EmbeddingGemma 2 unifies text, code, vision and audio in a single 768‑dim vector, delivering 14% higher code performance and 6× smaller storage.

Discover how OpenAI’s GPT‑6 uses looped transformers for deeper reasoning, impacting speed and model tiers. Understand the future of large‑scale AI inference.

Western open‑weights AI is rising, with Reflection AI’s Beam promising high‑efficiency coding and agentic performance as a Chinese‑free alternative.

Reddit users slam the new 64 GB DGX Spark for its high price and low value, sparking debate over NVIDIA’s pricing strategy and possible QC cuts.

New CivBench tests LLMs on long‑term strategy in Civilization V, revealing GLM‑5.3’s edge and Qwen‑3.8‑27B’s surprising science win—see how models stack up.

Explore the latest AI shifts: open‑weight models, small‑scale MoE training, llama.cpp performance tweaks, and new decision‑making benchmarks—insights that shape today’s industry.