AlephAlpha has released Kolibri, an open-weight Mixture-of-Experts model with 78.1B total parameters and 3.46B active parameters, designed for sovereign and regulated domains with a focus on German language performance.
The update reduces voice roughness and break-up while maintaining ~300 ms first-audio latency on laptop CPUs.
The MIT-licensed tool uses SQLite FTS5 and optional local embeddings to provide offline, Markdown-based memory for AI coding agents.
The 9B multimodal model claims 8 of 9 top scores in its size band on spatial benchmarks and includes an open-sourced data pipeline.
The agent, built on Anthropic's Claude Opus models, initiated contact with academics to solve coding bugs and validate its own work, sparking debate on AI autonomy and research ethics.
Adaptive EVS reduces VLM input tokens by 80% and improves concurrent stream throughput by 46%
The new DTOC framework allows agents to dynamically hide or unhide tool outputs, improving solve rates and reducing token costs on DeepSWE tasks.
A new paper co-authored by Geoffrey Hinton and Yoshua Bengio argues that automating AI research could compress years of progress into months, posing extreme control and societal risks.
LiquidAI has released LFM2.5-Encoder-350M and LFM2.5-Encoder-230M, with the 350M variant described as a multilingual bidirectional encoder supporting 15 languages and on-device inference.
The update includes 43 pull requests from 8 contributors, focusing on GPT-6.1 Sol integration, memory optimizations, and Windows stability fixes.
The 6B parameter model was trained on approximately 2,500 hours of real robot data and supports parallel grippers and dexterous hands.
Debian announced a self-service LLM inference portal on September 23, 2026, with recent updates adding the scaleway/qwen3.8-27b model and improved documentation.
Meta has open-sourced code allowing users to build their own Muse AI gadgets using hardware like ESP32 and Raspberry Pi, while also distributing 5,000 pre-built 'Muse Home Link' devices.
Anthropic is investing $100 million in the Claude Frontier Academy to train 10,000 'frontier deployed engineers' from partner companies by the end of 2027.
The 4B-parameter model is derived from Qwen3.5-4B and trained on 1,500 synthetic SWE tasks using the Leaf harness.
Muse has open-sourced its device SDKs and firmware under the Apache 2.0 license, allowing developers to build custom hardware interfaces for the Muse AI assistant using ESP32 and Linux devices.
GNOME CVE counts rose approximately tenfold from 2021 to 2026, driven largely by AI-generated reports, prompting a policy shift to accept AI-assisted findings while managing false positives and ending the bug bounty program.
Redis creator antirez has released ds4, a C-based inference engine supporting DeepSeek V4, GLM 5.x, and Qwen3.8 Flash Next on high-memory hardware.
Spotlight enables infinitely growing memory without increasing access costs by using an arbitrarily sparse indexing mechanism.
The program aims to train 10,000 Frontier Deployed Engineers by the end of 2027 to address the enterprise AI talent gap.
OpenAI has released a comprehensive model guide for the GPT-6 family, detailing strategies for managing context costs, selecting reasoning levels, and handling long-running workflows across GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna.
Streamline combines Cloudflare Workers, Containers, and Stream to enable real-time video manipulation like overlays and burned-in subtitles.
The new MIT-licensed benchmark evaluates models on finding and fixing real-world bugs in large codebases before users encounter them.
The platform allows agents to run on schedules or triggers with shared memory and requires human approval for risky actions.
AI Briefing