10 Hot AI Open‑Source Projects on GitHub This Week – The Last One Even Jensen Huang Praises

The AI open‑source community this week shows three clear trends: AI agents are moving from toy experiments to production‑grade tools, edge‑centric lightweight deployment is becoming mainstream, and major technology companies are accelerating open‑source releases across diverse domains.

1. Hermes Agent – A Self‑Evolving General‑Purpose AI Agent Framework

Hermes Agent, released by Nous Research under the MIT license, amassed 47 000 GitHub stars in six weeks, adding nearly 20 000 stars this week. Its standout feature is a "self‑improvement loop" that evaluates task outcomes, encodes successful experiences as reusable Skills, and enables cross‑session persistent memory, allowing agents to retain user profiles across conversations. The framework supports CLI, Telegram, Discord, Slack, WhatsApp and over 15 gateways, integrates more than 200 models via OpenRouter, and ships with 40+ ready‑to‑use tools. Implemented in Python, it offers tool invocation, a security sandbox, and scheduling mechanisms, directly addressing core pain points of long‑term collaborative agents. Tencent Cloud launched a Lighthouse lightweight‑application server template for one‑click cloud deployment of Hermes Agent.

Project address: https://github.com/NousResearch/hermes-agent

2. Evolver – Genome‑Evolution‑Protocol‑Based AI Agent Self‑Evolution Engine

Evolver, open‑sourced by EvoMap on 2026‑02‑01, recorded over 36 000 downloads in three days, now supports more than 130 000 AI agent nodes and 46 million cumulative calls. It applies genetic programming via a custom Genome Evolution Protocol (GEP) to let agents autonomously optimise prompts and action strategies, providing traceable, reusable evolution paths. Unlike Hermes Agent’s self‑improvement loop, Evolver manages a per‑agent "genome" that enables version rollback and knowledge transfer across agents, turning prompt optimisation into a systematic engineering process. Over 114 versions have been released, demonstrating high community activity. Recent controversy over alleged design copying by Hermes Agent underscores Evolver’s technical foresight.

Project address: https://github.com/EvoMap/evolver

3. OpenAI Agents SDK – Official Multi‑Agent Orchestration Framework

The openai‑agents‑python SDK, an official lightweight multi‑agent framework from OpenAI, has earned more than 22 000 stars. It provides a complete production‑grade toolchain for building AI agents, including multi‑agent orchestration, task delegation, tool calling, streaming, safety guards, context management, and deep integration with the latest OpenAI models. Compared with LangChain’s all‑inclusive approach, the SDK focuses on primitive primitives for agent collaboration, preserving lightness and flexibility.

Major updates on 2026‑04‑15 introduced a Harness mechanism for isolated testing and a SandboxAgent that mounts a containerised filesystem, enabling agents to reuse environments across requests for complex coding tasks. A forthcoming "sub‑agent" concept will allow primary agents to delegate specialised tasks to dedicated sub‑agents.

Project address: https://github.com/openai/openai-agents-python

4. Qwen3.6‑35B‑A3B – A Small‑But‑Mighty MoE Agent‑Programming Model

Alibaba’s Qwen3.6‑35B‑A3B, released on 2026‑04‑16, uses a sparse Mixture‑of‑Experts (MoE) architecture with 350 billion total parameters but activates only 30 billion per inference. Despite its smaller active size, it outperforms the dense 270 billion‑parameter Qwen3.5‑27B on several programming benchmarks and rivals Google’s Gemma4‑31B. The model also supports multimodal perception and reasoning, matching Claude Sonnet 4.5 on most vision‑language tasks and surpassing it on spatial‑intelligence benchmarks (RefCOCO 92.0, ODinW13 50.8). MoE’s “large total, small active” design reduces inference cost while preserving capability. The model can switch between "thinking" and "non‑thinking" modes and integrates with OpenClaw, Claude Code, and Qwen Code. Weights are hosted on Hugging Face and ModelScope for local deployment or Alibaba Cloud Bailei API access.

Project address: https://huggingface.co/Qwen/Qwen3.6-35B-A3B

5. HY‑World 2.0 – Tencent’s Multimodal 3D World Model

On 2026‑04‑16, Tencent open‑sourced HY‑World 2.0, a multimodal world model that ingests text, images, and video to generate, reconstruct, and simulate 3D environments. Its key breakthrough is editability: unlike Google’s Genie 3, HY‑World 2.0 outputs directly editable 3D assets (Mesh, 3DGS, point clouds) that can be exported to Unity, Unreal Engine, etc., enabling rapid creation of game‑ready maps and level prototypes from natural language or visual prompts. The framework unifies generation and reconstruction within an offline 3D world‑model paradigm and adds an interactive "role mode" for real‑time navigation with physics‑based collision. Applications extend beyond gaming to digital twins, architectural planning, and cultural heritage preservation. The accompanying paper is on arXiv, and code and weights are publicly released.

Project address: https://github.com/Tencent/HY-World

6. NVIDIA Ising – The World’s First Open‑Source Quantum‑AI Model Family

On 2026‑04‑15 NVIDIA unveiled the Ising family, targeting two core quantum‑computing challenges: error correction and processor calibration. The Ising‑Calibration model is a 350 billion‑parameter vision‑language model that interprets quantum processor measurements, reducing calibration time from days to hours. Trained on data from superconducting qubits, quantum dots, ions, and neutral atoms, it was co‑developed with Fermilab and Harvard. The benchmark QcalEval shows Ising‑Calibration‑1 achieving state‑of‑the‑art scores across six dimensions, surpassing Gemini 3.1 Pro, GPT 5.4, and Claude Opus 4.6.

The Ising‑Decoding model, built on a 3D CNN framework, accelerates quantum error‑correction decoding, delivering up to 2.5× speed‑up and 3× accuracy over the open‑source pyMatching baseline.

By turning AI into a "operating system" for quantum computers, Ising aims to lift the "5‑year curse" of quantum scaling, where error rates must drop to one in a trillion operations.

Project address: https://github.com/NVIDIA/ising

7. Omi – Real‑Time Screen‑Aware and Conversational AI Memory Assistant

Omi, from BasedHardware, is an MIT‑licensed multimodal AI assistant that watches the user’s screen, listens to conversation, transcribes speech, generates summaries and to‑do items, and retains everything in a continuous memory store. Its modular architecture separates hardware abstraction, AI inference, and application layers, supporting desktop, mobile, and wearables such as smart necklaces and AR glasses. With over 10 400 stars and 171 contributors, Omi demonstrates a shift from query‑based assistants to environment‑aware digital companions, offering an open‑source foundation for custom AI wearables.

Project address: https://github.com/BasedHardware/omi

8. Google AI Edge Gallery – Offline Mobile‑First AI Model Experience Platform

Google AI Edge Gallery, an open‑source Kotlin app, lets users run major LLMs (including Gemma 4) entirely offline on Android 12+ and iOS 17+ devices. It supports AI chat (with a "thinking" mode), camera‑based image analysis, audio transcription, and a 270 million‑parameter FunctionGemma model for on‑device operations. By keeping all inference local, the project eliminates network latency and privacy risks, offering a reference implementation for privacy‑sensitive domains such as finance, healthcare, and government.

Project address: https://github.com/google-ai-edge/gallery

9. ElatoAI – ESP32‑Based Real‑Time Voice AI Interaction via Edge Computing

ElatoAI combines an ESP32 microcontroller for audio capture/playback with Cloudflare edge functions for low‑latency session state and AI inference. Using OpenAI’s realtime API over secure WebSockets, the system streams audio to the cloud while the ESP32 handles only I/O. This architecture overcomes the typical edge‑AI bottleneck of insufficient compute on low‑cost hardware, providing a template for AI‑enabled toys, wearables, and voice assistants. The repository includes full embedded and server‑side code, illustrating end‑to‑end voice AI development.

Project address: https://github.com/akdeb/ElatoAI

10. OpenClaw – Zero‑Code, Local‑First AI Agent Automation Platform

OpenClaw (nicknamed "Little Lobster") is an MIT‑licensed, fully local AI agent framework that emphasises three differentiators: (1) all data and operations run on‑device, ensuring privacy for confidential workloads; (2) zero‑code deployment via a graphical interface enables setup within ten minutes even for non‑programmers; (3) comprehensive office‑automation capabilities covering file management, data processing, browser automation, and messaging, with built‑in features for file classification, spreadsheet generation, and information search.

The ecosystem now includes lightweight (Pico/NanoClaw), high‑performance (MaxClaw), and domain‑specific variants (MedClaw, ClawWork). Version v2.6.2 adds native video and music generation and a "dream" memory system that simulates human sleep. OpenClaw exemplifies the transition of AI from conversational tools to "digital executors" that let users harness large models without programming.

Project address: https://github.com/openclaw/openclaw

Summary and Trend Observations

Reviewing the ten projects reveals three major trends:

Trend 1: AI agents are evolving from toys to production tools. OpenAI Agents SDK’s SandboxAgent and Harness, Evolver’s GEP‑based traceable prompt optimisation, and Hermes Agent’s cross‑session memory all illustrate this shift.

Trend 2: Edge AI and lightweight deployment dominate. Google AI Edge Gallery runs large models offline on phones, Omi brings real‑time perception to desktops and wearables, and OpenClaw offers zero‑code local deployment, reflecting growing developer focus on on‑device security and efficiency.

Trend 3: Major vendors accelerate open‑source contributions. Tencent, Alibaba, NVIDIA, and OpenAI each released heavyweight projects within the same week, signalling that open‑source is now a strategic lever for technology diffusion and cross‑disciplinary innovation, especially as NVIDIA’s Ising brings AI into quantum computing.

For developers, the recommendation is to align project selection with specific business needs: use OpenAI Agents SDK or Evolver for complex AI systems, choose OpenClaw or Omi for privacy‑preserving local deployments, and explore NVIDIA Ising or HY‑World 2.0 for frontier research in quantum AI and 3D generation.