Granite Speech 5.0 is IBM's free, open speech model: transcription in six languages, fast on-device recognition and better meeting notes. What it does, and lacks.
Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, beating GPT-6 Astra on Moonshot's numbers. Here is how to use it now and what the open weights mean.
Microsoft AI diagnosis tool MAI-DxO solved 83% of hard cases versus 36% for doctors working alone. Here is how it works, the catches, and what comes next.
Recursive Flint is a free, Apache 2.0 coding model that runs offline on a laptop. What it does, how fast it is, its limits, and what it means for developers.
MiniMax M3.1 Flash is out on the MiniMax API at $0.15 per million input tokens and 200+ tokens a second. Here is what it does, what it costs and how it compares.
DeepSeek V4.1-Flash-1017 brings higher agent benchmark scores, open weights and a paid Search and Fetch API. Here is what changed and what it costs to use.
IBM Research has addressed the unreliability issue in artificial intelligence agents that pass tests on initial attempts but fail upon repetition. By evaluating decision uncertainty step by step, their new method identifies potential failure points and generates targeted guidelines to ensure repeatable execution.
Google has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to advance real time voice dialogue. These additions bring parallel background reasoning, mid conversation multilingual switching, and instant tool calling to conversational AI agents.
Every launch comes with a chart showing the new model winning. Here is what the chart leaves out, and which parts of the announcement are worth your attention.
A number that improves every quarter is reassuring and largely decorative. Measuring what a system is good at is much harder than ranking it.
Get a notification when something new is published.
Done. To stop them later, use the notification settings for this site in your browser.
