HySparse2 Slashes AI Agent Memory 4.5x While Cutting Prefill Work in Half
A research team introduces a two-level KV sharing scheme that cuts prefill compute by up to 5x while boosting long-context retrieval on 80B MoE models.
20 stories
A research team introduces a two-level KV sharing scheme that cuts prefill compute by up to 5x while boosting long-context retrieval on 80B MoE models.
A 3,500-star open source rules pack forces AI coding agents to strip out generic hero sections, invented metrics, and useless comments before shipping.
A Microsoft Research and UC Berkeley team shows agents sharing a scratchpad beat parallel independent runs, setting new records on ARC-AGI-3, polyomino packing, and MNIST compression.
Ant Group's Ling team open sourced five Ming 3.0 speech models under MIT, with the 0.5B TTS topping open-weight voice benchmarks.
NetEase Youdao's open-source R2T2 streams speech to text with 200-600ms latency and never rewrites words it already emitted.
Moonshot rebrands Kimi WebBridge as the Kimi Browser Extension, adding a sidebar chat, recordable skills, and cross-agent support that runs locally.
Lightricks released a free IC-LoRA that turns raw Blender or Unreal clay animations into finished cinematic shots while locking camera and object positions.
Tencent Hunyuan's new interactive benchmark actually runs AI-generated web apps, uses code coverage to hunt broken paths, and matches human taste 85.3% of the time.
A Chinese quant repo reproducing 100+ brokerage research strategies is trending, bundling RSRS timing, factor models, and HHT algorithms into runnable notebooks.
An Android accessibility-service app reads WeChat, QQ, and X conversations, then drafts three ranked replies you can paste with one tap.
A Codex Skill that turns raw copy into a six-shot stickman video plan with locked prompts for Gemini Omni Flash, no API required.
A new research paper proposes decoupling style decisions from waveform generation, using explicit instructions and preference optimization to keep voices consistent across turns.
Viggle's v0.2 LoRA compresses Qwen-Image-2.1 from 40 diffusion steps to 5, keeping sample diversity at 93% of the teacher.
Moonshot's 2.8T-parameter open-weight model lands on AWS with a 1M-token context, native vision, and explicit prompt caching for agent workloads.
Tencent's Hunyuan team launched a preview of Hy Image 3.5, claiming a 30% win rate boost over 3.0 at $0.024 per image.
A new benchmark scores text to speech models on how correctly they pronounce tricky words, and voice preference does not predict accuracy.
StepFun's new 600B MoE flagship matches Kimi K3 on the Intelligence Index at roughly one third the cost, with open weights due in October.
A new paper introduces Hill Sampling, a minimalist test-time search that beats AlphaEvolve and evolutionary weight training on hard math problems.
xAI's Grok 4.7 jumps 111 Elo on Artificial Analysis's agentic knowledge work benchmark, closing the gap with Anthropic at roughly half the cost per task.
Cognition ships xAI's newest model in Devin CLI and Desktop, where it excels at hard backend work but over-scopes on simpler tasks.