MirroS launched Code-as-World: a paradigm that represents bodily worlds by means of executable world representations. The argument is slender and testable: pixels are proof of a bodily scene, not its ontology. A video mannequin can predict believable frames with out ever representing mass, contact, or gravity. So as an alternative of pixels, latents, or captions,…
Google Research and UNSW Sydney have launched GlucoFM, a self-supervised basis mannequin for continuous glucose monitoring. Its core transfer is a cut up. Existing CGM fashions — CGMformer, GluFormer, CGM-JEPA — encode a glucose hint as one entangled sequence. GlucoFM decomposes it right into a sluggish physiological “state” stream and a transient “occasion” stream, retains…
Alibaba’s Qwen workforce has launched Qwen3.8-Flash-Next, an open-weight multimodal Mixture-of-Experts mannequin constructed for price per token. The checkpoint pairs a 125B spine with a 51B N-gram embedding desk and a 4B multi-token prediction module. Only 6B parameters activate per token. The workforce positions it as an early preview of the structure that may underpin Qwen4,…
IBM has launched Granite 4.2, a household of open reasoning language fashions in 3B, 8B, and 30B parameter sizes. Unlike earlier Granite releases, which have been instruction-following assistants, Granite 4.2 is constructed round specific reasoning. Every mannequin can emit a sequence of thought earlier than answering, and each mannequin exposes a considering / non-thinking swap…
Model playing cards report high quality below server-class, full-precision circumstances. Those numbers hardly ever predict how the identical mannequin behaves on a cellphone. This week, Liquid AI launched Pipette. It is an open-source platform for benchmarking basis fashions on edge gadgets, inbuilt partnership with Artificial Analysis as an unbiased methodology validator. Pipette treats on-device habits…
Perplexity has launched Portable Computer, a local-first construct of its agentic Computer platform that runs the agent harness, orchestrator, planner, software router and post-trained fashions immediately on NVIDIA DGX Spark. The native mannequin, inference engine, software sandbox and app connectors ship as one packaged system, each job begins on the gadget, and work dealt with…
Training and serving frontier fashions is now a networking drawback as a lot as a compute drawback. Collective operations like all-reduce and all-to-all synchronize hundreds of accelerators throughout coaching, and the slowest switch units the tempo for your entire job. Even small quantities of community friction instantly strand important compute capability. This week, Meta launched…
Information extraction groups face a recurring selection. Small encoder fashions are low-cost however inflexible, and huge language fashions are versatile however costly per doc. Fastino launched GLiNER2.5 to slim that hole. The launch replaces span enumeration with boundary prediction: the mannequin scores the place an entity begins and ends as a substitute of scoring each…
A workforce from Google Research and USC has launched Mobility-Embedded POIs (ME-POIs), a framework that folds combination human motion into text-based place embeddings. The premise is that language fashions describe what a place is, however not how it’s used. Two espresso retailers can share a class, an handle block, and a textual content vector, whereas…
Frontier open-weight fashions are transport quicker than the {hardware} assumptions round them. Kimi-K3, GLM-5.2 and DeepSeek-V4-Flash are closing the aptitude hole with proprietary methods, however releasing parameters solely determines who can get hold of a mannequin — not who can afford to run it. Serving them nonetheless assumes datacenter-class GPU clusters, and as agentic workloads…