AI — Page 22

A 2,500-task training set raised the model's score on AutomationBench-AA's public split from 34.0% to 57.7% in about 11 hours.

The six-model release pairs Arabic language models with speech recognition and publishes weights, training checkpoints and evaluation code under Apache 2.0.

The San Francisco startup says custom ROCm kernels and lower-precision math cut a 2K edit from about 22 seconds on the same MI355X hardware.

The test used simulated company workloads; Poetiq says its custom routers improved cost and accuracy against OpenRouter and TypeSafe's JEV on its benchmark.

The October 15th update adds voices for French, German, Hindi, Mexican Spanish, Korean and Japanese; OpenAI says the API version will follow.

Comet 60B keeps text, image and video inputs, with downloadable weights in three formats; Multiverse's benchmark results are company-reported.

FLUX 3 Image Klein adds editing and multi-reference generation, with weights for fine-tuning and an API priced from $0.009 per 1K images.

Isaiah Granet's latest product push automates website and API setup, while Bland's existing iMessage offering is still described as an enterprise service.

CEO Ramin Hasani's MIT-born startup says the model leads its 112-test suite; it returns probabilities for text and image decisions without generating output tokens.

The arrangement puts CEO Dario Amodei's September call for embedded oversight into practice, with publication rights and defined limits on what evaluators can see.

Alexandr Wang's team is releasing the weights for coding and long-horizon tasks, with local operation requiring two M5 Ultra Mac Studios or substantial GPU memory.

Nemotron 3.5 Super VL combines text, image, document and video processing; NVIDIA says it completed 10,000 coding tasks up to four times faster in its tests.

The model debuted on September 30th with access limited to cyber defenders; Google opened its Gemini app and AI Mode rollout on October 13th.

Step 5 Preview has 600 billion total parameters and 27 billion active per token. NVIDIA says its NVFP4 version runs at nearly 8,000 tokens per second per GPU on GB300.

TBC's first product runs on conventional chips; the company says it makes video generation faster and cheaper, claims that remain unverified by independent testing.

The collaboration pairs insitro founder Daphne Koller's biology and machine-learning platform with AstraZeneca's drug-development expertise to pursue MASH and cardiomyopathy treatments.

Founder Oleksandr Kosovan is extending MacPaw's Mac-utility business into AI task automation. Eney asks before consequential actions by default, while its AI requests are processed in the cloud.

CEO and co-founder Arvind Raichur says patients can ask plain-language questions about their records. wisepatientAI described natural-language search in October 2025, and its App Store version history lists the feature in January 2025.

CEO Li Dahai's team says expert preloading makes the open model practical on mobile; the weights still rely on device storage, and the speed figures are company-reported.

Its new map links 29 benchmarks to Labor Department occupations, but Mercor says the coverage estimate is provisional and depends on its selection rules.

Founder Tyler Johnston's October 14th video revisits investigations alleging links between anonymous X accounts and Leading the Future, whose PAC reported $31 million in cash at June's end.

FILM6 organized the October 7th-8th event; the jury's judging process and the festival's claimed 1,100 submissions remain unaccounted for in public records.

Runway is adding isolated deployments and fine-tuning to enterprise access for its video models.

Anthropic's confirmation warns that other feedback on the same conversation must also be withdrawn before it leaves the improvement program.