The Open-Source AI War Just Exploded: April 2026 Scorecard

In one week in early April 2026, six AI labs dropped six major open-source model releases. The models are impressive. The real story is the license shift.

Five of six models are now Apache 2.0 — fully permissive, enterprise-safe, no custom restrictions. Open source is no longer a compromise. It's a competitive advantage.

The April 2026 Leaderboard

Why This Week Changes Everything

Google moved Gemma to Apache 2.0.

No more custom license friction for enterprise. Gemma 4 26B MoE achieves 88.3% on AIME 2026 with only 3.8B active parameters — runs on a single 24GB GPU. That's Codex-tier performance on consumer hardware.

Qwen 3.6 Plus went free on OpenRouter.

1M token context. Free preview for a limited time. Chinese labs (Alibaba, Zhipu AI) are now leading on specific benchmarks, with GLM-5 trained entirely on Huawei chips with zero NVIDIA dependency.

Every release is pushing toward agents, not chat.

Native function calling, long context windows, efficient MoE architectures — all the characteristics that make models useful for agentic workflows, not just conversational ones.

OpenAI released its first open model.

gpt-oss-120B under Apache 2.0 is a significant strategic shift. The company that defined proprietary frontier AI is now open-sourcing at the 120B scale.

The License Shift Is the Real Story

5 of 6 models are now fully permissive. The implications:

• Enterprise can deploy without legal review of custom licenses
• Derivative models can be fine-tuned and redistributed freely
• Competitive moats based on model access are eroding
• The gap between frontier and free is narrowing fast

A free Chinese model is now finding real bugs in production codebases with Codex-tier performance. A year ago, that sentence would have seemed implausible.

Model Selection Decision Framework

For April 2026, here's how to choose:

• Edge/Mobile: Gemma 4 E2B (phone, 4GB RAM) or E4B (laptop, 8GB)
• Long context (1M+): Qwen 3.6 Plus — MoE, 1M tokens, agentic coding
• Maximum reasoning: Gemma 4 31B — needs H100 (80GB) but highest benchmark scores
• Cost efficiency: Gemma 4 26B MoE — 88.3% AIME on a single 24GB GPU

The connecting thread: every model in this release batch is optimizing for agents, not chat. Function calling. Long context. Efficient inference for agentic loops. The direction is clear.

The Bottom Line

A year ago, AI coding meant autocomplete. Now it means deploying autonomous agents across repos while a free model finds real bugs in your code.

Open source is no longer a compromise. The question isn't whether to use open models — it's which one, for which task, at what context length.

Jason Newell is an AI practitioner, builder, and writer covering agentic systems, developer tooling, and the future of AI engineering.