April 2026 delivered one of the densest months of AI model releases in recent memory. Seven major open source models launched in just the first twelve days, from Meta's massive Llama 4 family to Google's phone-sized Gemma 3n. Here's what shipped, how they compare, and which ones deserve your attention.
The Full Lineup
Seven models launched between April 2–9, spanning dense architectures, mixture-of-experts (MoE), code generation, and on-device multimodal. Here's the breakdown.
| Model | Org | Parameters (Total / Active) | Architecture | License | Standout Feature |
|---|---|---|---|---|---|
| Llama 4 Scout | Meta | 109B / 17B | MoE (16 experts) | Llama 4 Community | 10M token context window |
| OLMo 2 32B | Ai2 | 32B / 32B | Dense | Apache 2.0 | Fully open training data and code |
| Llama 4 Maverick | Meta | 400B / 17B | MoE (128 experts) | Llama 4 Community | Best multilingual MoE performance |
| Qwen 3 72B | Alibaba | 72B / 72B | Dense | Apache 2.0 | Top dense model on reasoning tasks |
| Qwen 3 MoE 235B | Alibaba | 235B / 22B | MoE | Apache 2.0 | Near-frontier at low active params |
| Codestral 2 | Mistral | 22B / 22B | Dense | Apache 2.0 | Code generation, fill-in-the-middle |
| Gemma 3n | 4B effective / 2B footprint | Dense multimodal | Gemma License | Runs on-device (phone, tablet) |
The Big Stories
Llama 4: MoE Goes Mainstream
Meta released two MoE models simultaneously, and both are impressive for different reasons.
Llama 4 Scout is the practical one. With a 10M token context window (yes, ten million), it handles document analysis and long-context tasks that would choke other models. At only 17B active parameters, it's efficient enough to run on a multi-GPU consumer setup while still hitting near-frontier benchmarks.
Llama 4 Maverick is the powerhouse. 400B total parameters across 128 experts, but only 17B active at inference. It delivers the best multilingual MoE performance seen in open source, making it ideal for global applications.
Both use the Llama 4 Community License, which is permissive for most commercial uses but worth reviewing if you're building at scale.
Qwen 3: Alibaba's Dense Champion
Qwen 3 72B quietly became the top-performing dense model on reasoning benchmarks. For tasks that need consistent, high-quality outputs without MoE complexity, this is the model to watch. The Apache 2.0 license makes it freely usable commercially.
The MoE variant (Qwen 3 MoE 235B) hits near-frontier quality while keeping active parameters at just 22B — a sweet spot for anyone balancing cost and performance.
Gemma 3n: AI in Your Pocket
Google's Gemma 3n takes the opposite approach from the big models. At 4B effective parameters with a 2B footprint, it's designed to run entirely on phones and tablets. It handles text, images, and audio locally — no cloud required.
For privacy-first applications and on-device experiences, this opens possibilities that weren't practical before.
Codestral 2: Code Gets Its Own Model
Mistral's Codestral 2 is a focused 22B model built for code generation and fill-in-the-middle tasks. Apache 2.0 licensed and specifically tuned for programming, it's a strong option for developer tools and coding assistants.
OLMo 2 32B: The Fully Open Option
Ai2's OLMo 2 32B is the most transparent model in the lineup. Training data, code, and methodology are all fully open under Apache 2.0. If reproducibility and auditability matter to your use case, this is your model.
What This Means for You
For developers building products: The Llama 4 and Qwen 3 families give you near-frontier performance at a fraction of the cost of proprietary APIs. The MoE architectures mean you can run impressive models on reasonable hardware.
For teams needing on-device AI: Gemma 3n is a breakthrough. Multimodal inference on a phone without network calls changes what mobile AI apps can do.
For anyone who values openness: OLMo 2 32B and the Qwen 3 models (Apache 2.0) give you truly open alternatives with no licensing ambiguity.
For code-focused workflows: Codestral 2 fills the gap between general-purpose models and specialized code tools.
The Bigger Picture
This month marks a clear shift in the open source AI landscape. MoE architectures are no longer experimental — they're production-ready and shipping from major labs. The gap between proprietary and open source models continues to narrow, and the licensing is getting more permissive, not less.
The tools and frameworks ecosystem is maturing too. Google's ADK for multi-agent orchestration, Meta's Llama Stack for unified deployment, and OpenAI's Codex CLI for terminal-based coding all launched alongside these models. The infrastructure is catching up to the capabilities.
Which Model Should You Use?
- Long-context document analysis: Llama 4 Scout
- Multilingual applications: Llama 4 Maverick
- Best general reasoning (dense): Qwen 3 72B
- Cost-efficient near-frontier: Qwen 3 MoE 235B
- Mobile/on-device: Gemma 3n
- Code generation: Codestral 2
- Maximum transparency: OLMo 2 32B
The pace isn't slowing down. If April is any indication, 2026 is going to be the year open source AI became the default choice for most applications.
Want to compare these models side by side? Check out the Q4KM AI Model Directory for detailed specs, benchmarks, and pricing across every major model.