Trending archive for Oct 21, 2026

deepseek-ai/DeepInfer

DeepSeek Opens Its Production Inference Engine — DeepInfer is the offline inference framework DeepSeek uses internally for large-scale serving, now open-sourced for DeepSeek-V4.1-Flash. It bundles optimized kernels (FlashMLA, DeepGEMM, DeepEP), MoE/CP parallelism, speculative decoding with confidence-scheduled verification, and batch-invariant/deterministic sampling—useful for teams deploying DeepSeek models at scale who need reproducible outputs.

DeepInfer: High-performance, large-scale inference framework for DeepSeek models

fyzugx/DEsk

A Wayland desktop that renders your screen as a literal desk — DEsk is a Linux Wayland desktop environment using realistic skeuomorphic design, like iOS 6: windows are paper, the dock is a drawer, apps sit in binders, notifications arrive by post box. Now at v0.1.2 beta with installers for Debian, Ubuntu, Arch and Fedora, it's for Linux users wanting a playful, tactile alternative to flat desktop UIs.

jyjblrd/single-propeller-drone

A drone that flies with just one spinning propeller — This project documents an under-actuated drone stabilized by a single propeller mounted atop a finned body, instead of the usual multi-rotor setup. It includes custom forks of ArduPilot and moteus for flight and motor control, plus Fusion 360/STEP files for the body and propeller designs, useful for hobbyists building unconventional drones.

Under-actuated drone with only one motor

tywr/Nordwand-Mono

A Python-built monospace font blending Univers with JetBrains Mono proportions — Nordwand-Mono is a coding font designed from scratch in Python, merging the classic Univers neo-grotesque with JetBrains Mono's compact proportions. It offers alternative character styles, stylistic sets for ligatures and wide-angled brackets, and lets you compile a customized version by editing a config file. Good fit for developers who want a hackable, Swiss-inspired terminal font.

A compact monospace neo-grotesque font designed for developers (and adventurers).

togethercomputer/HybridAgent

An LLM agent that switches between latent thinking and text actions — HybridAgent lets an LLM agent alternate between latent-space reasoning and text-based tool use within one generation pass, using a HybridLink module to map hidden states back to embeddings. It ships HybridSFT/HybridRL training stages, batch inference, and a tool to render latent thoughts as readable pseudo-text for interpretability.

google-research/harness2

Self-improving agent scaffolding without touching model weights — Harness² is a Google Research framework that recursively refines the prompts, tools, and orchestration around a fixed model. It uses a Propose-Probe-Compose loop to adapt the harness per task, plus a persistent improvement harness that carries lessons across tasks. Aimed at researchers building agents for open-world, unseen tasks.

receptron/mulmoclaude

A local-first AI assistant you grow with plain-file memories and no-code apps — MulmoClaude lets Claude build and maintain a personal wiki, schema-driven data, and small single-user apps (lists, trackers, flashcards) entirely as local files — no programming needed. Everything stays in your workspace under MIT, and you can reach the same assistant from your phone or a messaging app.

Nurture your own AI assistant on your own computer. Local-first and MIT: memories, data and apps stay as plain files in your workspace. Chat summons the right GUI — wiki, spreadsheet, chart, form, 3D. Build small apps for an audience of one, no programming required.

receptron/mulmoterminal

A Browser Grid for Juggling Several Coding Agents at Once — MulmoTerminal is a local, tmux-backed browser terminal that runs multiple coding-agent CLIs (Claude Code, Codex, Grok, Copilot CLI, Cursor CLI and more) side by side, colour-coding each cell as working, done, or waiting on you. Sessions survive reloads, isolate via git worktrees, and can push phone alerts when a turn finishes — built for people juggling several agents, not just one.

Run multiple Claude Code and Codex sessions in parallel — a browser terminal grid that shows which agent needs you. Local, tmux-backed, MIT.

musoukun/skills

A grab-bag of drop-in skills for Claude Code and similar AI agents — This repo collects reusable skill folders for Claude Code-style AI agents, each with a SKILL.md you copy into your local skills directory. Included skills cover annotating screenshots for guides, DDD-style requirements modeling, generating narrated 3D explainer videos with Three.js, pattern/case classification, and turning architecture diagrams into animated MP4/GIFs.

cloudflare/security-audit-skill

Turning coding agents into structured, self-verifying security auditors — This skill orchestrates isolated agents through a six-phase security audit: recon, coverage-led hunting, candidate validation, structured findings, independent verification, and reporting. It's the single-repo seed behind Cloudflare's larger vulnerability discovery harness, useful for teams wanting reproducible, machine-readable audit trails rather than one-shot AI scans.

A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings

ebichan88/dbxray

Turn a live database into readable Markdown docs and ER diagrams — dbxray connects to a database and generates table lists, per-table definition docs, and Mermaid ER diagrams in Markdown, viewable directly on GitHub. It also exposes this data through an MCP server so AI assistants can answer questions about table structures, making it useful for both documentation and AI-assisted schema exploration.

dbxrayは、DBに接続し、テーブル一覧・各テーブルの定義書・ER図などをMarkdown形式で出力するツールです。また、出力したデータをMCPサーバー経由でAIに回答することもできます。