Alibaba's HydraHead Cuts Transformer Memory 65% by Mixing Attention Heads
Researchers from Alibaba Group show that hybridizing attention at the head level — guided by mechanistic interpretability — unlocks 512K context with only 15B training tokens
20 stories
Researchers from Alibaba Group show that hybridizing attention at the head level — guided by mechanistic interpretability — unlocks 512K context with only 15B training tokens

H2O LLM Studio is an open-source, no-code GUI that lets anyone fine-tune large language models without writing a single line of code.

ClaudePrism brings Claude, LaTeX, and Python into a single offline-first desktop app for academic writing — a local alternative to OpenAI's cloud-based Prism

NVIDIA's FP4-quantized Qwen3.5-397B fits on 4 Blackwell GPUs instead of 8, with near-zero accuracy loss across every benchmark tested.
Gemma 4 31B hits Cerebras public preview at 1,800+ tokens/sec, marking the platform's first multimodal model and first Google DeepMind collaboration

Runway partners with MIXI to deploy AI video tools across Japan's gaming giant, cutting production timelines from weeks to days

Google's TabFM does zero-shot tabular classification and regression in one forward pass, no training or hyperparameter tuning required.

Intelligent Internet's open-source Zenith harness lifts GPT-5.5 from 5th to 1st on FrontierSWE, beating Claude Fable without needing a bigger model

Git 2.55 ships incremental MIDX repacking, a new git history fixup command, Linux fsmonitor support, and a raft of performance improvements

Replit launches a native desktop app for Mac and Windows, bringing distraction-free AI-powered development outside the browser
Google AI Studio's Build mode now lets you bundle your full chat history when sharing an app, making prompt workflows instantly reproducible for anyone with the link.

Claude Opus 4.8 and Haiku 4.5 graduate from preview to GA in Microsoft Foundry, unlocking Azure-native billing, MACC drawdown, and a new 'hosted on Azure' mode with US data residency

Cursor launches a native iOS app that lets developers spawn and supervise AI coding agents from anywhere, as the company prepares to close its $60B SpaceX acquisition

Cognition's Devin Fusion uses a parallel 'sidekick' agent and dynamic mid-session routing to cut frontier-model costs by 35% without sacrificing code quality.

Meta's Brain2Qwerty v2 decodes full sentences from brain waves without surgery, hitting 61% word accuracy and releasing all training code publicly
Can GLM-5.2 finally replace closed models for agentic coding?

Tencent's HunyuanVideo has grown from a single 13B text-to-video model into a full open-source ecosystem rivaling closed-source leaders like Runway and Sora

ChatDev 2.0 (DevAll) transforms from a software-only agent system into a zero-code multi-agent platform for any complex task, backed by NeurIPS 2025 research.

AEON-7 abliterates Ornith-1.0-35B with zero capability loss, then squeezes 3x faster decode on a DGX Spark via NVFP4 + DFlash speculative decoding.

vLLM-Omni squeezes up to 172% more audio throughput from four TTS models by treating each pipeline stage as a separate optimization problem