Models
Frontier and open models: what shipped, what it can do, and how the claims hold up.
-
Google pitches Gemini 4 Argon as a robot brain, with its own benchmark score
Google says its frontier Argon model scores 87% on its own embodied reasoning benchmark and can generate the simulation data needed to…
-
A 0.88B world-action model runs robot policies on edge hardware
A small research group has cut NVIDIA’s two-billion-parameter Cosmos Policy down to 0.88 billion parameters and reports it producing robot action chunks…
-
OpenAI begins watermarking its models’ text output
OpenAI opened invisible text watermarking to API customers and published its detection rates, with an EU rollout to follow.
-
Abu Dhabi’s TII releases an Arabic-first speech recognition model
Falcon ASR is a 1.6bn-parameter speech model built around Arabic and the Emirati dialect, with word error rates TII says beat the…
-
Nadella argues frontier models should be treated as insider risks
Microsoft’s chief executive published a design argument at the weekend for governing frontier models the way enterprises govern powerful insiders, including an…
-
FAR.AI reports internal probes catch model sabotage that text monitoring misses
FAR.AI reports that probes reading a model’s internal activations beat text-based monitoring at spotting sabotage, on the authors’ own benchmark numbers.
-
Xiaomi publishes MiMo-V2.6 and the recipe for scaling its reinforcement learning
Xiaomi’s new MiMo-V2.6 report is a reinforcement learning scaling recipe, with open weights under an MIT licence and, unusually, no benchmark headline…
-
Sakana AI’s Japanese-tuned model goes into a clinical evidence search tool
Sakana Namazu is now behind Evidence Finder, a doctor-facing literature search tool, with a 96.4% medical licensing exam score and a very…
-
Abu Dhabi’s TII releases a speech model aimed at Emirati Arabic
TII’s Falcon-ASR reports the best published error rate on an Arabic leaderboard, while its Emirati dialect result rests on the institute’s own…
-
LightOn ships three OCR models and picks apart its own benchmarks
LightOn has released OCR models at three sizes and spends much of the announcement explaining why the edit-distance scores it reports are…