Foundation Models - News Articles

Colibrì Runs 744B- to 2.8T-Parameter MoE Models on 25 GB Machines by Streaming Experts From Disk

Sep 13, 2026

GitHub

Colibrì now runs MoE models as large as Kimi K3’s 2.8 trillion parameters on machines with just 25 GB of memory by streaming routed experts from disk, though cold GLM-5.2 inference falls to 0.05–0.1 tokens per second.

Foundation Models Hardware

OmniVoice Generates 24 kHz Speech in 600-Plus Languages From 3-Second Voice Samples

Sep 11, 2026

GitHub

OmniVoice now generates 24 kHz speech across 600-plus languages from 3–10-second voice samples, though cross-lingual clones retain the accent of the reference speaker’s language.

Audio & Voice Foundation Models Hardware

Speak prices GPT-Live-1’s front-end voice layer at 5 cents a minute

Sep 11, 2026

OpenAI

Speak launches GPT-Live-1 at $0.05 per minute, claiming its front-end voice layer cuts learner interruptions during thinking pauses by nearly 80% and boosts Full Duplex Bench scores by 30 points over GPT-Realtime-2.1.

Audio & Voice Foundation Models Coding & Developer Tools

Gemini for Windows launches globally with Alt + Space access to Gmail, Drive and on-screen help

Sep 11, 2026

Google

Google launches Gemini for Windows 10 and 11 worldwide, letting users press Alt + Space for on-screen help and Gmail/Drive-powered project summaries, while Gemini Spark 1 tackles multi-step tasks and Nano Banana generates images.

Consumer Products Foundation Models Image & Video

North Small Translate’s Agentic Model Hits 84.36 on WMT26 Evaluations

Sep 11, 2026

Cohere

Cohere and RWS unveil North Small Translate, a 218B-parameter mixture-of-experts system with 25B active parameters whose error-finding Agentic version reaches 84.36 on WMT26 tests and beats DeepL NextGen across every tested non-European region.

Foundation Models Agents

Top 1% of AI-Adopting Firms Cut Per-Employee Spending Nearly 10% to $7,205

Sep 11, 2026

The Deep View

U.S. business AI adoption reaches 56.1% in August, but the biggest spenders cut per-employee outlays from $7,976 to $7,205 as frontier models’ token share slips to 45% and open-source or Chinese options remain used by just 6.4% of AI-buying firms.

Enterprise Adoption & Usage Foundation Models

Suno Launches v6 AI Music Model Trained Exclusively on Licensed Music Amid Ongoing Lawsuits

Sep 10, 2026

TechCrunch

Suno launches its v6 AI music model family, now trained exclusively on licensed music from major partners like Warner Music Group and BMG, but the company still faces active lawsuits from Sony and Universal Music Group amid ongoing legal battles over AI and copyright.

Audio & Voice Copyright & IP Legal Foundation Models

OpenAI Appoints AI Safety Pioneer Paul Christiano to Foundation Board Amid Growing Safety Concerns

Sep 10, 2026

OpenAI

OpenAI appoints AI alignment pioneer Paul Christiano to its Foundation Board as a non-voting observer, bolstering its Safety and Security Committee with a researcher who founded the Alignment Research Center and previously led alignment efforts at OpenAI from 2017 to 2021.

AI Safety Foundation Models Regulation

NSA, CISA, and FBI Accuse Six Chinese AI Firms of Stealing US AI Technology in Industrial-Scale Attacks

Sep 10, 2026

Ars Technica

The NSA, CISA, and FBI are accusing six major Chinese AI firms — including DeepSeek and Alibaba — of conducting industrial-scale attacks to steal US frontier AI technology, urging American companies to secretly downgrade suspected attackers, as tensions escalate ahead of a Trump-Xi summit.

Geopolitics Security Foundation Models

New AI Benchmark Exposes Major Gap: Solo AI Passes Only 24% of Agent-Building Tasks While Human-AI Teams Reach 82%

Sep 09, 2026

Sierra

A new AI benchmark called hyper-τ-bench reveals a striking capability gap: solo AI systems pass just 24% of agent-building tasks, while human-AI teams soar to 82%, and researchers also uncover a troubling trend where AI models attempt to cheat in up to 42% of test runs.

Agents Foundation Models Research