Categories
AI News iRaluca

Inception’s Mercury 3 Drafts Every Word at Once, 1,000 Tokens a Second

Inception’s Mercury 3 is a diffusion LLM that refines whole answers in parallel: 1,068 tokens/s and 1M context. The benchmarks are its own.

Categories
AI News iRaluca

GLM-5.3 Asked Itself If It Was Training Data, Then Chose to Behave

Apollo Research caught Z.ai’s open-weight GLM-5.3 reasoning, unprompted, about how its actions might shape its future self. Rare, but a first.

Categories
AI News iRaluca

Anthropic Drafted a Bill on When Its Own Models Could Be Stopped

Anthropic has quietly pitched Congress a draft AI safety bill. It sets when Washington could halt a frontier model, and who gets to decide.

Categories
AI News

Beam’s Weights Are Finally Out. Reflection’s Real Pitch Is Shorter Thinking

Reflection’s open-weight Beam is out under Apache 2.0. Its pitch: near-frontier reasoning in far fewer tokens. Independent tests mostly agree.

Categories
AI News

DeepSeek’s Flash Update Came With a Search Engine for Other AIs to Read

DeepSeek’s minor V4.1-Flash update brought open weights and a coding jump, plus a bigger surprise: its own search engine, sold to AI agents.

Categories
AI News iRaluca

Nvidia’s Bolt Fits a 74B Model in 24GB by Shrinking Its Experts to 2 Bits

Nvidia’s open Nemotron 3.5 Bolt squeezes a 120B parent into 74B parameters and 24GB, with 2-bit experts. What the model card shows, and what it skips.

Categories
AI News iRaluca

Microsoft Grew a 2.8-Trillion-Parameter Model. Nobody Can Talk to It Yet

Microsoft finished pre-training MAI-Base-2, a 2.8T-parameter model it says is 4–12x more efficient than DeepSeek’s best base model. Not usable yet.

Categories
AI News iRaluca

Apertus 1.5 Shipped Its Weights in July. This Week It Shipped the Recipe

Swiss AI’s Apertus 1.5 adds images, audio and a 262k context. Now every training stage, dataset and script behind it is public too.

Categories
AI News iRaluca

Meta’s Open Muse Is Hard to Jailbreak. Fine-Tuning Could Change That

Meta’s open-weight Muse Spark Open barely jailbreaks in its own tests, but refuses about 1 in 6 harmless prompts. And fine-tuning is allowed.

Categories
AI News

How Many Words You Hide Decides What UltraBERT Is Good At

Open encoder family UltraBERT trains with two masking ratios at once to be good at both classifying text and finding it. Apache 2.0.