Open Source AI Projects Announcements: Week of April 5, 2026

The first week of April 2026 was one of the densest periods for open source AI releases in recent memory. Six major labs shipped competitive open-weight models within days of each other, an agent framework went viral overnight, and Block donated its AI coding agent to the Linux Foundation. If you blinked, you missed at least three launches.

Here is everything that shipped, what it means, and which projects are worth your attention.

The Big Picture: Open Source Hits Parity

For years the narrative was simple: closed models lead, open models follow 6 to 12 months behind. That story broke in early 2026. The models released this week match or beat proprietary alternatives on key benchmarks, and several run on consumer hardware.

| Project | Org | Parameters | License | Key Benchmark | Runs Locally | |---|---|---|---|---|---| | Gemma 4 (31B Dense) | Google DeepMind | 31B | Apache 2.0 | 89.2% AIME 2026 | Yes (single GPU) | | GLM-5.1 | Zhipu AI (Z.ai) | 744B MoE / 40B active | MIT | #1 SWE-Bench Pro | FP8 quantized | | Llama 4 Scout | Meta | 109B MoE / 17B active | Meta Community | 10M token context | Yes | | Llama 4 Maverick | Meta | 400B MoE / 17B active | Meta Community | 85.5% MMLU | Needs multi-GPU | | Mistral Small 4 | Mistral AI | 119B MoE / 6B active | Apache 2.0 | 40% faster than v3 | Yes (llama.cpp) | | Goose | Block / Linux Foundation | N/A (agent framework) | Apache 2.0 | N/A | Yes (Rust) | | Claw Code | Community | N/A (agent framework) | Open source | 72k stars in 48h | Yes |

The common thread: mixture-of-experts architectures that keep active parameter counts low enough for local inference while total parameter counts push into the hundreds of billions.

Gemma 4: Google Goes Apache 2.0

Google DeepMind released Gemma 4 on April 2 with four model sizes: E2B, E4B, 26B MoE, and 31B Dense. This is the first Gemma generation with a fully OSI-approved Apache 2.0 license, which matters for commercial use.

The numbers that stand out:

• 256K context window across all sizes
• Multimodal input (text and image on all models, audio on edge variants)
• 140+ languages supported
• The 31B Dense model scores 89.2% on AIME 2026 and 80.0% on LiveCodeBench v6, beating models with 3x the parameters

The E2B and E4B edge models run on a Raspberry Pi. Day-one support covers Hugging Face transformers, GGUF, ONNX, vLLM, and Ollama. If you are building a local-first AI application on macOS or Linux, Gemma 4 is now a serious option for on-device inference.

GLM-5.1: A 744B Model Trained Entirely on Huawei Ascend

Zhipu AI shipped GLM-5.1 on April 7, a post-training upgrade to GLM-5. The headline: a 744B parameter MoE model with 40B active parameters per token, trained entirely on Huawei Ascend chips with zero NVIDIA hardware.

The base GLM-5 scored 50.4% on Humanity's Last Exam. GLM-5.1 focuses on coding and ranked #1 on SWE-Bench Pro at time of release. Released under MIT license, weights available on Hugging Face in full precision and FP8 quantized.

What makes this notable beyond the benchmark numbers is the hardware story. Training a competitive 744B model without any NVIDIA GPUs demonstrates that the Ascend ecosystem is production-ready for frontier-scale training, not just inference.

Meta Llama 4: Scout and Maverick

Meta released two new open-weight models in early April under the Meta Community License:

• Llama 4 Scout: 109B total parameters, 17B active, with a 10 million token context window. Ten million. That is not a typo.
• Llama 4 Maverick: 400B total, 17B active, 128-expert MoE architecture scoring 85.5% on MMLU

The 10M context on Scout is the largest publicly available context window in any open model. For applications like codebase analysis, long document processing, or agent memory, this changes what is possible without chunking or retrieval augmentation.

Mistral Small 4: Three Products in One

Mistral Small 4 shipped on March 16 and saw heavy adoption through early April. At 119B parameters with 128 experts (6B active per token), it unifies three previously separate Mistral products into one model:

1. Magistral (reasoning)
2. Pixtral (multimodal vision)
3. Devstral (agentic coding)

It ships with a configurable reasoning effort toggle, 256K context, and native text + image input. Mistral claims a 40% reduction in end-to-end completion time versus their previous generation. Available under Apache 2.0 on Hugging Face, vLLM, and llama.cpp.

Goose: Block's Agent Joins the Linux Foundation

Goose is an open source AI agent from Block (Jack Dorsey's company) that was donated to the Agentic AI Foundation at the Linux Foundation on April 7. Built in Rust, it goes beyond code suggestions: it installs, executes, edits, and tests code directly.

Key details:

• Model agnostic: works with 15+ providers (Anthropic, OpenAI, Google, Ollama, Azure, Bedrock)
• Native desktop app for macOS, Linux, and Windows, plus CLI and embeddable API
• General purpose: handles research, writing, automation, and data analysis alongside code

The Linux Foundation donation signals that enterprises are taking open agentic infrastructure seriously as a category, not just a collection of side projects.

Claw Code: The Viral Agent Framework

The wildest story of the week. On March 31, a security researcher discovered that Claude Code's complete source code (~512,000 lines of TypeScript) was accidentally exposed via an npm source map. Developer Sigrid Jin created Claw Code, a clean-room Python and Rust rewrite of the agent harness architecture, and open-sourced it. It hit 72,000 GitHub stars and 72,600 forks in its first 48 hours, making it one of the fastest-growing repos in AI tooling history.

Claw Code focuses on the "harness layer" that connects LLMs to tools, file systems, and workflows. Whether you view this as a cautionary tale about source map hygiene or a validation of the agent harness pattern, the community response shows massive demand for open alternatives to proprietary coding agents.

Architecture Overview: How These Projects Fit Together

The announcements this week fall into two categories: foundation models (the LLMs themselves) and agent frameworks (the harness that makes them useful). Here is how they relate:

Agent frameworks sit on top and are model-agnostic. You can swap the foundation model underneath without changing your application code. This is the pattern that makes the open source stack composable: Goose works with Gemma 4 through Ollama just as well as it works with Claude through the Anthropic API.

What This Means for Developers Building AI Agents

If you are building AI agents that run on user machines (desktop automation, local assistants, privacy-sensitive workflows), this week changed the calculus:

How to Pick the Right Model for Your Use Case

| Use Case | Best Fit | Why | |---|---|---| | On-device agent (macOS/Linux) | Gemma 4 E4B or 26B MoE | Runs locally with Ollama, Apache 2.0 | | Long document processing | Llama 4 Scout | 10M token context, no chunking needed | | Code generation and review | GLM-5.1 | #1 SWE-Bench Pro, MIT license | | Multimodal (text + image) | Mistral Small 4 | Unified instruct + vision + code | | General chat / assistant | Llama 4 Maverick | 85.5% MMLU, strong all-rounder | | Desktop automation agent | Fazm + any model via Ollama | macOS-native, model-agnostic |

Common Pitfalls When Adopting New Open Source Models

• Benchmark gaming. A model that scores 89% on AIME might still hallucinate on your specific domain. Always test on your actual workload before switching from a model you know works.
• Quantization trade-offs. FP8 and GGUF quantized versions are smaller and faster but lose accuracy on edge cases. If your agent makes tool calls that need to be exactly right (file paths, shell commands), test quantized models specifically on those patterns.
• License confusion. "Open source" and "open weight" are not the same thing. Llama 4 is open-weight under the Meta Community License, which restricts commercial use above certain thresholds. If you are building a product, read the license text, not the marketing blog post.
• Context window != usable context. A 10M token window does not mean the model performs equally well at token 9,999,999 as it does at token 1,000. Needle-in-a-haystack performance degrades with length. Benchmark retrieval at the context sizes you actually plan to use.

Quick Reference: Links and Resources

| Project | GitHub / Hugging Face | License | |---|---|---| | Gemma 4 | google/gemma-4 | Apache 2.0 | | GLM-5.1 | zai-org/GLM-5 | MIT | | Llama 4 Scout | meta-llama | Meta Community | | Llama 4 Maverick | meta-llama | Meta Community | | Mistral Small 4 | mistralai | Apache 2.0 | | Goose | block/goose | Apache 2.0 | | Claw Code | claw-code.codes | Open source |

Wrapping Up

The week of April 5, 2026 demonstrated that open source AI is no longer playing catch-up. Multiple competitive models shipped under permissive licenses, agent frameworks matured into Linux Foundation projects, and the tooling to run these locally on consumer hardware arrived alongside the models themselves. If you are building desktop AI agents, the foundation model layer is now a solved problem; the differentiation is in the agent harness, perception, and action layers on top.

Fazm is an open source macOS AI agent that works with any of these models. Open source on GitHub.