Open Source AI Projects Releases and Announcements: April 2026

April 2026 has delivered the densest stretch of open source AI activity since the field went mainstream. Frontier-class models shipped under permissive licenses, agent frameworks joined major foundations, inference engines added day-one support for new architectures, and governance decisions reshaped how the ecosystem coordinates. This post covers every significant release and announcement in one place, organized so you can find what matters for your work.

April 2026 Releases and Announcements at a Glance

| Project | Date | Type | Category | License | |---|---|---|---|---| | Qwen 3 (0.6B to 235B) | April 5 | Release | Foundation model | Apache 2.0 | | Gemma 4 (2B, 12B, 27B) | April 9 | Release | Foundation model | Gemma License | | GLM-5.1 (744B MoE) | April 7 | Release | Foundation model | MIT | | Llama 4 Scout/Maverick | April 5 | Release | Foundation model | Llama 4 License | | Mistral Small 4 (24B) | April 8 | Release | Foundation model | Apache 2.0 | | MiniMax M2.7 | April 12 | Announcement | Self-evolving model | Open source | | Goose joins Linux Foundation | April 8 | Announcement | Agent framework | Apache 2.0 | | MCP under AAIF governance | April 10 | Announcement | Protocol standard | Open governance | | Google ADK | April 9 | Release | Agent framework | Apache 2.0 | | OpenAI Agents SDK | April 7 | Release | Agent framework | MIT | | vLLM 0.8 | April 6 | Release | Inference engine | Apache 2.0 | | Shopify AI Toolkit | April 10 | Announcement | Developer tools | MIT | | Overworld Waypoint-1.5 | April 11 | Announcement | 3D generation | Open source | | OpenAI Codex CLI update | April 12 | Release | Developer tools | MIT | | ComfyUI v0.19 | April 12 | Release | Image generation | GPL-3.0 |

Foundation Model Releases

Qwen 3: Eight Sizes, One Architecture

Alibaba's Qwen team shipped Qwen 3 on April 5 with eight model sizes ranging from 0.6B to 235B parameters. The headline release is the 235B Mixture-of-Experts variant (235B total, 22B active), which competes with GPT-4.5 and Claude Opus on reasoning benchmarks while running on substantially less hardware than its parameter count suggests.

All Qwen 3 models ship under Apache 2.0, meaning no usage restrictions for commercial or research applications. The dense models (0.6B through 32B) are designed to run on consumer hardware, while the MoE variants target multi-GPU workstation and server deployments.

| Model | Parameters | Active Params | Target Hardware | Standout Capability | |---|---|---|---|---| | Qwen3-235B-A22B | 235B MoE | 22B | Multi-GPU server | Top reasoning scores, Apache 2.0 | | Qwen3-32B | 32B dense | 32B | Single A100/H100 | Best dense model at size | | Qwen3-14B | 14B dense | 14B | RTX 4090 / M3 Max | Strong coding, fits consumer GPU | | Qwen3-7B | 7B dense | 7B | RTX 3060+ / M2 Pro | Fast local inference | | Qwen3-4B | 4B dense | 4B | RTX 2060+ / M1 | Mobile-class deployment | | Qwen3-1.7B | 1.7B dense | 1.7B | Any GPU / CPU | Edge and embedded use | | Qwen3-0.6B | 0.6B dense | 0.6B | CPU only | Smallest, fastest |

Qwen 3 introduced a "thinking mode" toggle across all sizes. You can switch between extended reasoning (like chain-of-thought) and direct response at inference time, which is useful for applications where you want reasoning for some queries but fast responses for others.

Gemma 4: Google DeepMind's Open Weights Play

Google DeepMind released Gemma 4 on April 9 in three sizes: 2B, 12B, and 27B parameters. The 27B model matches the performance of much larger models from prior generations on several benchmarks, particularly in multilingual and multimodal tasks.

Gemma 4 ships with built-in tool use capabilities, meaning the model can natively generate structured function calls without special prompting tricks. This is significant for agent applications where the model needs to interact with external APIs.

The llama.cpp community had GGUF quantizations available within hours of release, and Ollama added support the same day.

GLM-5.1: Zhipu AI's 744B MoE Under MIT

Zhipu AI's GLM-5.1 is the largest model released this month at 744B parameters (MoE). It shipped under MIT license on April 7 and immediately drew attention for its top score on SWE-Bench Pro, a benchmark that tests a model's ability to resolve real GitHub issues.

The community response was rapid. By April 11, Unsloth had dynamic 2-bit quantizations reducing the model from 1.4TB to 220GB, and multiple GGUF variants appeared on Hugging Face for different hardware configurations.

Llama 4 Scout and Maverick

Meta released Llama 4 in two variants: Scout (17B active / 109B total MoE) and Maverick (17B active / 400B total MoE). Scout targets efficient local deployment with a 10M token context window, while Maverick targets quality-critical applications.

Mistral Small 4: 24B That Runs Anywhere

Mistral Small 4 (24B parameters, Apache 2.0) shipped on April 8 with an emphasis on running efficiently on consumer hardware. The model fits in 16GB of VRAM at 4-bit quantization, making it one of the strongest models you can run on a single consumer GPU or an M-series Mac.

MiniMax M2.7: Self-Evolving Agent Model

MiniMax announced and open-sourced M2.7 on April 12, a "self-evolving" model designed specifically for agentic workflows. The model can update its own system prompts and tool configurations based on task outcomes, a capability aimed at long-running autonomous agent deployments.

Agent Framework Announcements

Goose Joins the Linux Foundation

Block's AI agent framework Goose officially joined the Linux Foundation on April 8 under the Agentic AI Foundation (AAIF). This moves Goose from a corporate-sponsored project to community governance, with contributions from multiple organizations. Goose focuses on local-first execution where your code and data never leave your machine.

MCP Under Open Governance

The Model Context Protocol (MCP), originally developed at Anthropic, moved under AAIF governance in early April. This is one of the most consequential announcements of the month because MCP is becoming the standard way AI agents interact with tools and data sources. Open governance means no single company controls the protocol's evolution.

Google Agent Development Kit (ADK)

Google released ADK on April 9, an open source framework for building AI agents that integrates with Gemini models and supports MCP out of the box. ADK provides session management, memory, and tool orchestration in a single package.

OpenAI Agents SDK

OpenAI shipped its Agents SDK on April 7 under MIT license. The SDK provides primitives for building multi-agent systems with handoffs, guardrails, and tool execution. It works with OpenAI's models and supports the Responses API.

Inference and Infrastructure

vLLM 0.8

vLLM 0.8 shipped on April 6 with FP8/FP4 quantization support, improved speculative decoding, and better multi-GPU scaling. The release included day-one support for serving Qwen 3 and GLM-5.1.

llama.cpp Updates

The llama.cpp project shipped multiple updates through the month, adding support for Gemma 4 on release day and continuing to optimize quantization for MoE architectures. Build system improvements on April 12-13 (b8779) cleaned up CMake configurations and reduced build times.

Ollama

Ollama released version 0.20.6 on April 12 with updated model support and performance improvements. The update included better memory management for large MoE models and faster model loading times.

Developer Tooling

Shopify AI Toolkit

Shopify announced its AI Toolkit on April 10 with MCP plugins for Claude Code, Cursor, and Gemini CLI. The toolkit provides structured access to Shopify's APIs through the MCP protocol, so AI coding assistants can read and modify Shopify store configurations directly.

OpenAI Codex CLI

The Codex CLI received a significant update on April 12 with stronger sandbox networking controls. The update hardens the execution environment for running AI-generated code, addressing concerns about code execution safety in agent workflows.

ComfyUI v0.19

ComfyUI shipped version 0.19 on April 12 with workflow improvements and new node types for image generation pipelines. The release focuses on stability and performance rather than new features.

Announcements That Shape the Ecosystem

Several announcements this month matter not because they shipped code, but because they changed the rules.

Apache 2.0 as the default. Qwen 3, Mistral Small 4, Google ADK, and Goose all shipped under Apache 2.0 or MIT. The trend toward maximally permissive licensing for open source AI accelerated in April, making it easier to build commercial products on top of these models without legal ambiguity.

MoE as the dominant architecture. Four of the five major model releases this month (Qwen 3 235B, GLM-5.1, Llama 4, and Gemma 4 27B internally) use Mixture-of-Experts. This architecture keeps the total parameter count high (for knowledge capacity) while keeping the active parameter count low (for inference speed), and it is now clearly the preferred approach for frontier open models.

Local-first agent execution. Goose, Codex CLI, and the broader MCP ecosystem all emphasize running agents locally rather than through cloud APIs. Combined with models like Mistral Small 4 and Qwen3-14B that fit on consumer hardware, the infrastructure for fully local AI agent workflows is now in place.

What to Watch Next

The pace of open source AI releases in April 2026 shows no signs of slowing. Key areas to monitor:

• Qwen 3 fine-tuning ecosystem. With eight model sizes under Apache 2.0, expect a wave of domain-specific fine-tunes through the rest of April and into May.
• GLM-5.1 accessibility. Community quantization work continues to push the model toward more affordable hardware configurations.
• MCP adoption. As the protocol moves under open governance, more tools and services will add MCP support, expanding what AI agents can interact with.
• Agent framework consolidation. With Goose, ADK, OpenAI Agents SDK, and several others all active, the next few months will show which frameworks attract sustained developer adoption.

Open source AI in April 2026 is not just about model releases. The combination of permissive licensing, open governance, local-first tooling, and community-driven optimization is building an ecosystem where anyone can deploy capable AI systems without depending on a single provider. The announcements this month make that future tangible.