
Qwen3.6 Shrinks to 3.4 GB After Cerebras REAP Deletes 90% of Experts
A community experiment uses REAP expert pruning to shrink Qwen3.6-35B down to 6B params and 3.4 GB, trading coherence for extreme portability
20 stories

A community experiment uses REAP expert pruning to shrink Qwen3.6-35B down to 6B params and 3.4 GB, trading coherence for extreme portability

Deep-Live-Cam hits 94k GitHub stars with real-time face swapping from a single photo, no training required, running live on consumer GPUs
Code2LoRA uses a hypernetwork to instantly generate repo-specific LoRA adapters, matching per-repo fine-tuning accuracy with zero inference-time token cost

An open-source engine from AIDC-AI turns a single topic prompt into a complete short video with script, visuals, voiceover, and music.

A deliberately broken Qwen3 MoE model passes 15/15 validation checks after 90% of its experts are deleted, exposing a systemic flaw in how compressed models get evaluated.

Aperant (formerly Auto Claude) brings a Kanban-driven multi-agent pipeline to your local desktop, running up to 12 parallel Claude agents in isolated git worktrees to plan, build, test, and ship software autonomously.
FreeLLMAPI is a self-hosted TypeScript proxy that stacks free-tier keys from 16 AI providers behind one OpenAI-compatible endpoint, adding up to ~1.7B tokens per month with automatic failover.
Browser Use's web automation library lands as a native Claude Code plugin, letting agents browse, click, and scrape without leaving your coding session

Google's open-source Visual Blocks lets you wire up ML pipelines by dragging nodes — no code required, and now with LLM-powered pipeline generation via InstructPipe.
earthtojake/text-to-cad gives any coding agent a full CAD pipeline locally — parametric Python output, STEP/STL/URDF exports, and a browser viewer, all MIT-licensed and free

FrancyJGLisboa's open-source repo turns any workflow description into a validated, cross-platform agent skill — no spec writing or prompt engineering needed
Suno's revamped stem separation now regenerates individual tracks from scratch, delivering artifact-free audio ready for professional DAW workflows.
Obscura is a Rust-built headless browser that uses 6x less memory than Chrome and ships with built-in stealth mode, targeting AI agent workloads at scale
Databricks open-sources Omnigent, a meta-harness that unifies Claude Code, Codex, and custom agents under one shared layer for composition, control, and real-time collaboration.
Warp's AI terminal agent now connects directly to SuperGrok subscriptions, unlocking Grok Build 0.1 — xAI's purpose-built agentic coding model — without burning Warp credits.

gpu-hot is a self-hosted, Docker-based NVIDIA GPU dashboard that delivers sub-second metrics and scales from a single workstation to 100+ node clusters
LLM Wiki turns Karpathy's viral knowledge-base pattern into a full desktop app — compiling your documents into a persistent, interlinked wiki instead of re-answering from scratch every time.
Z.ai's GLM-5.2 arrives as the new coding flagship with 1M-context support, two thinking-effort modes, and a full open-source MIT release coming next week
OmniVoice Studio is a free, fully local desktop app that clones voices, dubs videos, and transcribes speech across 646 languages — no API keys, no subscription, no cloud.
ViMax turns any idea, script, or novel into a multi-scene video by coordinating a team of specialized AI agents — solving the consistency and length limits that have plagued AI video generation.