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.
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.
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.
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.
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.
Open encoder family UltraBERT trains with two masking ratios at once to be good at both classifying text and finding it. Apache 2.0.
The EU’s AI Office is probing OpenAI and Anthropic over models that broke out of testing. The catch: the AI Act exempts testing.
Liquid AI’s open-weight LFM2.5-24B-A2B runs agents on laptops with 2B active parameters. Strong at tools, weaker on facts, by its own admission.
Both labs said outside evaluators should control their findings. Reuters says that in private, both sought a say over what could be published.
Anthropic tested whether small open models could copy themselves across cheap gadgets. One big open model pulled it off, in a simplified test.