Google DeepMind — Gemma 4
Released April 2, 2026 under Apache 2.0 — the most permissive open license in Gemma history.
Four Variants:
| Model | Parameters | Use Case |
|---|---|---|
| Gemma 4 2B | 2B | Phones, edge devices |
| Gemma 4 4B | 4B | Phones, edge devices |
| Gemma 4 26B | 26B MoE | High-performance |
| Gemma 4 31B | 31B Dense | #3 globally on Arena AI |
Key Specs:
- All 4 support text, images, video
- Edge models handle native audio input
- Context windows: 256,000 tokens
- 31B scores 89.2% on AIME 2026
- Arena Elo: 1452
Day-one support: Hugging Face, Ollama, vLLM, llama.cpp, MLX, LM Studio, NVIDIA NIM, Android Studio
Microsoft — MAI Trio
Launched April 2, 2026 through Microsoft Foundry.
MAI-Transcribe-1
- 3.8% Word Error Rate (WER) across 25 languages
- Beats Whisper Large v3 on all 25
- Beats Gemini 3.1 Flash on 22/25
MAI-Voice-1
- 60x real-time audio generation
- Custom voice from seconds of sample audio
- $22 per million characters
- Competes directly with ElevenLabs
MAI-Image-2
- Debuted in Arena.ai top 3
- 2x faster than predecessor
- $5 per million tokens input
- $33 per million tokens output
- WPP as first enterprise partner
Built by a team of 10 people — CEO Mustafa Suleiman’s philosophy of small, empowered teams.
Microsoft Copilot — Critique & Model Council
Two new orchestration features:
| Feature | What it does |
|---|---|
| Critique | One model generates, second reviews for accuracy |
| Model Council | Side-by-side comparisons of multiple models |