
Korean Startup Motif Technologies Ships Motif 3, Beating Trillion-Parameter Giants on Agent Tasks
Motif Technologies released Motif 3, a 314B parameter sparse MoE with a novel GDLA attention, native 256K context, and MIT license.
20 stories

Motif Technologies released Motif 3, a 314B parameter sparse MoE with a novel GDLA attention, native 256K context, and MIT license.
Unsloth's GGUF quants shrink MiniMax H3, a 20B omni-modal video model with native stereo audio, down to sizes that fit on a single consumer GPU.

An abliterated MLX build of DeepSeek V4 Flash squeezes a 284B MoE into 95GB, running unrestricted at 100k context on a 128GB Mac.

Anthropic retrained Fable 5's biology classifier, cutting biology-related fallbacks by 85% while keeping dual-use research locked behind trusted access pathways.

Beads gives coding agents a structured, versioned task graph so they stop forgetting what they were doing mid-project

pi-subagents adds a code mode that lets you orchestrate Pi child agents with plain JavaScript — loops, fan-out, branching, and git worktree isolation all in one script

Codex Security Review now automatically scans GitHub PRs for vulnerabilities using full repo context, threat models, and sandbox validation — free during preview.

Moonshot AI's 2.8-trillion-parameter open-weight model lands on Databricks with enterprise governance, slashing costs 50-72% vs. proprietary alternatives

Google's Gemini 3.6 Flash hits 60.4% on ARC-AGI-2 at just $0.61 per task, while Flash-Lite offers a budget-friendly 10.3% at $0.14

Epoch AI's new Mystery Game Puzzles benchmark hides its game identity to stop models from gaming the test -- Opus 5 leads at 59%

Browser Use Cloud now lets AI agents pay for browser sessions with USDC via Coinbase's x402 protocol, eliminating the need for API keys or account signup.

AMD acquires Toronto startup Taalas, which hardwires AI models directly into silicon, promising 17,000 tokens/sec at 20x lower cost than GPUs

Artificial Analysis patches its Intelligence Index to v4.1.1, upgrading grader models and fixing scoring errors in the τ³-Banking benchmark

Hugging Face's Candle brings PyTorch-like ML to Rust, enabling serverless inference and Python-free production deployments with 20k+ GitHub stars

OpenAI upgrades GPT-5.6 Sol for paid ChatGPT users with 68% fewer factual errors and gives free users unlimited chats with GPT-5.6 Luna plus a new Think button

Reka open-sources 10,312 hours of unscripted, first-person household video under Apache 2.0, giving physical AI researchers a large-scale real-world training corpus
Wispr Flow turns messy speech into clean, context-aware text across your IDE, terminal, and AI coding tools.

OpenAI, Microsoft, Cursor, AWS, and Vercel launch Agent Plugins v1.0: one plugin format that works across Codex, ChatGPT, Copilot, VS Code, Cursor, and Kiro.

Google DeepMind's WeatherNext achieves a decade of forecasting progress in one leap, now open-sourced with 1,000-scenario ensemble predictions per storm

Meta's AI models swept five STEM Olympiads — including perfect scores in physics — with no tools, no calculators, and no search allowed.