Home

Open Source

Curated open-source AI projects with verified GitHub data and hands-on breakdowns.

Ant Ming-Image: MIT is free for commercial use, Qwen is not

A fact-check of inclusionAI/Ming-Image (2026-09-24 GitHub API snapshot: 91 stars, 6 forks, Python, MIT). One design line ships two weights: Design for text-to-design, producing UI, dashboards, infographics and posters end to end, and Design-Layer, which splits a flat design into 2 to 9 semantically independent RGBA transparent layers. The architecture is a 6.15B design transformer plus a 17.01B multimodal LLM plus a 3.09B connector, about 26.44B total (vendor basis). It offers a native RGBA VAE, 8K structured prompts, and sampling defaults of steps 12 and CFG 1.0 for Design (2.0 for Layer), with recommended resolutions 2048 and 1024. The license red line: MIT allows commercial use, self-hosting and derivative work, contrasting with Qwen-Image 2.1's non-commercial Qwen Research License. The hardware floor is a single 80 GiB GPU at BF16; speed and benchmark figures are tagged vendor basis and not independently retested.

2,007 stars on your phone, but no free commercial use

omnimind-ai/OmniBot is an on-device AI agent that runs directly on Android handsets, built with native Kotlin and Flutter plus a React and TypeScript WebUI, and aimed at a full loop of understanding, deciding, acting and reporting back rather than being another chat shell. A 2026-09-23 GitHub snapshot shows 2,007 stars, 144 forks, Dart as the primary language, created 2026-03-18, last pushed 2026-09-23 and 27 open issues, figures that are a same-day cross-section and change daily. Its capabilities cover Skills, an Alpine environment that runs Linux on the phone, a browser, MCP and Android system-level tools, plus scheduled tasks, alarms, calendar and audio control, short and long-term memory, and file, workspace, browser and terminal access. Three technical highlights stand out: on-device unity, where the app runs on the phone without a computer as host; a real command-line workspace from Alpine, the terminal and ReTerminal; and the Remote Codex bridge, which after running the bridge on a computer lets the phone scan a code to borrow that machine's Codex execution. Onboarding means configuring an AI provider and scene models in settings, where Memory embedding strictly requires an embedding model and other scenes are best served by multimodal or vision models, with a local service defaulting to port 8899. The licence section carries the most weight: it uses segmented dual licensing, under which AGPL v3 is free only for non-commercial or personal, educational and research use and explicitly bars use by any organization, while commercial use, or avoiding AGPL's source-disclosure duty, requires signing a commercial licence first. The article fixes three conclusions: it cannot be called simply AGPL open source, it cannot be described as free for commercial use, and organizations are outside the free scope even for research. It also notes that GitHub's licence field reports NOASSERTION and cannot be trusted, so the findings come from the LICENSE file itself, and it draws a clear boundary with the site's earlier phone-harness piece, since that one is an iPhone driven by a Mac harness while this is Android-native on-device, differing in platform, architecture and licence.

Skip the agent loop: 12,078 stars for an open AI workspace

krillinai/OpenCreator is an open-source AI creation workbench and Skills collection maintained by the krillinai team (formerly KrillinAI), released under the permissive Apache-2.0 license and usable commercially and self-hosted. As of a 2026-09-22 GitHub snapshot it shows 12,078 stars, 1,222 forks, TypeScript as primary language, created 2024-12-17, last pushed 2026-09-21 and 31 open issues, figures that are a same-day snapshot rather than a lasting level. Its defining design choice is local-first: project data, attachments and logs stay on your machine by default (SQLite and the file system, with Codex sessions and config in CODEX_HOME), the Daemon listens only on 127.0.0.1 with Bearer tokens required for every endpoint except health checks, HTML preview disables scripts and navigation by default, and the desktop package enables ASAR integrity checks and Cookie encryption. The pivotal architectural judgment is that it does not rebuild the Agent loop but uses Codex CLI directly as the execution engine, wrapping it with three layers, a local Runtime, a visual workbench and a desktop host, so the agent loop, sessions, reasoning, tool calls, Skills and MCP all come from Codex; the upside is not maintaining a second engine, inheriting Codex's capabilities and using Codex-native configuration for Skills and MCP, while the cost is a hard dependency on the Codex ecosystem, a capability ceiling set by Codex, and available models depending on your local Codex and AI service settings. The README's prose says ten creation tools while its own table lists twelve rows, ten available and two in development (Auto Clips and Digital Avatar), plus seven built-in video-production Skills (KrillinAI CLI, Subtitle, TTS, Landscape and Portrait Render, Cover, Pipeline Plan). It is worth stating that a Skill in the repository is not auto-installed nor are external services bundled, and that it is not an open-source replacement for JianYing since its core is Agent plus creation tools plus Skill orchestration without a multi-track timeline editor.

French-to-English Only, 1,519 Stars: Why Hibiki Matters in 2026

Kyutai's kyutai-labs/hibiki is an open streaming speech translation model (1,519 stars, 119 forks, primary language Rust, created 2025-02-04, last push 2026-09-09, 11 open issues, per the 2026-09-21 GitHub API) that reuses Moshi's multistream architecture. It is decoder-only and models source and target speech jointly, emitting text and audio tokens at a constant 12.5Hz; the 2B variant uses 16 RVQ channels per stream and the 1B uses 8 for on-device use, trained on sequences up to 120 seconds with a 40-second inference context, described in arXiv 2502.03382. Its real highlight is training: word-aligned same-speaker data does not exist at scale, so the team uses contextual alignment, a weakly supervised method leveraging the off-the-shelf MADLAD translation system for word matching, under the rule that a word appears in the target only once predictable from the source, applied by inserting silences or synthesizing with a voice-controlled, alignment-aware TTS. Inference relies only on simple temperature sampling and is therefore compatible with batching, while voice fidelity is tuned through the CFG coefficient, default 1 and typically 3, with excessive values hurting translation. Limits are stated plainly: French to English only today, weights under CC-BY 4.0 requiring attribution, code split MIT for Python and the web client plus Apache-2.0 for the Rust backend, and a core implementation shared with Moshi so both repositories must be read.

MiniMax Opens Its Deck: mcode, the Terminal Agent You Can Audit

MiniMax open-sourced mcode, its terminal coding agent: repository MiniMax-AI/minimax-code (1,443 stars, 159 forks, TypeScript, MIT, created 2026-06-01, last push 2026-09-20, per the 2026-09-20 GitHub API), pitched as continuously unlocking model capability through excellent harness design. The core claim: the coding-agent battlefield has moved from the model to the harness, where permissions, sandboxing and auditability decide whether enterprises dare to use it. Three entry points (interactive TUI, headless mcode exec, ACP), BYOK to OpenAI and Anthropic compatible APIs, plus MCP, skills, parallel subagents and AGENTS.md. The vendor reports a 76.7 percent FrontierHarness pass rate at a 4 minute 33 second median. A cold look: 1,443 stars is still early and plugin-ecosystem depth is unproven, but for regulated industries auditability can outweigh a few points of pass rate.

Qwen-MM-Plugins Deep Dive: Multimodal Plugins for Any Agent

QwenLM/Qwen-MM-Plugins (2,908 stars, Python, Apache-2.0, created 2026-07-29, last push 2026-09-18, per the 2026-09-19 GitHub API) positions itself as "make any agent harness multimodal-native": a Skill-plus-MCP layered set of on-demand perception plugins that plugs into Claude Code, OpenClaw and other mainstream harnesses, fixing the 2026 gap that coding agents cannot watch or listen. The core claim: backed by the official QwenLM ecosystem and co-evolved with Qwen3.8-Omni-Flash, it is the concrete move in the model-plus-toolchain play. Apache-2.0 carries no commercial red lines, but the plugins are deeply tied to Qwen-family models, so budget for migration cost when switching backends.

Context Mode: Taming the Coding-Agent Context Window

mksglu/context-mode (23,324 stars, TypeScript, Elastic License 2.0, created 2026-02-23, last push 2026-09-16, per the 2026-09-18 GitHub API) positions itself as "context window optimization for AI coding agents": an MCP-layer sandbox intercepts and compacts context, backed by a SQLite/FTS5 knowledge base and session-continuity design across 17 clients. The core claim: it hits three real pains, bloated long sessions, diluted key instructions, and token cost that grows with length. But be honest about the license: ELv2 is not OSI-approved open source and carries two red lines, no offering it as a hosted service and no removing the license notice, fine for personal use but a legal review before corporate adoption.

herdr Deep Dive: The Runtime Layer Your Coding Agents Live On

herdrdev/herdr (39,133 stars, Rust, Apache-2.0, created 2026-03-27, ranked No.8 on the OpenGithubs 20260914 weekly with 2,458 stars added) positions itself as "the runtime your coding agents live on": work detaches but keeps running, multiple machines share one window, every pane is marked working, blocked or idle, agents drive it through the CLI and socket API, and it ships as one Rust binary with no Electron. The core claim: it claims the seat tmux would occupy in the AI era, and the agent-native socket API is what separates it from a tmux wrapper. But the project is under half a year old with unsettled APIs and storage formats, so manage your dev machine and experimental agents first, not your production-critical path.

Intern-S2: 397B multimodal base with pluggable memory

In September 2026 Shanghai AI Lab fully open-sourced Intern-S2, a 397B MoE multimodal foundation model: code at github.com/InternLM/Intern-S1, weights at HuggingFace internlm/Intern-S2-397B, also on ModelScope. Its core Memory Decoder introduces a pluggable external memory module that decouples knowledge storage from reasoning, so switching domains needs no base retraining; the Mobius architecture lifts end-to-end inference efficiency nearly 4x. Per the lab's own reporting, general ability ranks among the top open-source models, it leads on scientific long-horizon tasks such as Biology-Instructions, Mol-Instructions and MP20, and matches Gemini 3.1 Pro on IMO-Proof and AdvancedMathBench. This piece notes plainly that 397B is impractical to self-host, and most benchmarks are lab-reported with limited independent replication.

diagram-design: AI diagrams as deliverable static files

The GitHub repo cathrynlavery/diagram-design ranked second on the OpenGithubs weekly momentum chart dated 2026-09-14, gaining 7,208 stars that week; verified on 2026-09-15 it holds 39,807 stars, 2,528 forks, HTML as its main language, an MIT license, created 2026-04-16, last pushed 2026-09-10, with only 44 open issues. It is a diagram skill pack for Agent Skills compatible hosts including Claude Code, Codex, Factory Droid, Pi, GitHub Copilot, Kiro and OpenCode, and the official README claims 39 editorial diagram types, while the weekly chart blurb says 38, a discrepancy this piece resolves in favor of the README. Its output is self-contained HTML with inline SVG: no build step, no JavaScript, no external image dependency, openable offline by double-click, with each type shipping three static variants, minimal light, minimal dark and full-editorial. The design system is what defeats the AI look: a single accent color, one or two focal elements per diagram, 1px hairline borders, no shadows, a 10px border-radius ceiling, and every coordinate and gap divisible by four. It can redraw draw.io, Mermaid and Excalidraw sources into that system through four dials, format, size, detail and audience, emitting a fidelity ledger; it inherits components, relationships, grouping and direction but never source coordinates, palette or fonts. Its tagline is No Mermaid slop, yet it ships a Mermaid import path, a tension worth reading closely. The piece also covers brand onboarding that reads your homepage for palette and font stack, maps them to semantic tokens like paper, ink, muted and accent, checks WCAG AA contrast and emits a fidelity receipt; multi-client profile isolation; and the genuinely serious engineering: CI across three platforms, clipping detected by pixel diffing rather than geometry, plus gates for Sankey conservation, waterfall running totals, treemap area error and label collision, all built to catch diagrams that lie.