Gemma 4
Gemma 4 (also tracked as Gemma 3 or Gemma4) is a language model first observed on March 9, 2026, with a hosted price of $0 per million input tokens and $0 per million output tokens. As of March 2026, no public technical report, model card, or official statement has confirmed the parameter count, architecture, training data, or license. The presence of multiple naming variants has created ambiguity over whether Gemma 3 and Gemma 4 represent distinct releases or a single model cataloged under inconsistent identifiers. This zero-cost structure eliminates per-token pricing as a barrier to experimentation and sets a new baseline for free-tier language model access. This entity is notable because its appearance introduces an unverified, free-access model into the developer ecosystem at a moment when provenance, safety evaluation, and licensing transparency are standard prerequisites for enterprise and research adoption, making standard due diligence impossible at the time of its initial detection.
Google's Gemma 4 has become its fastest-launching model ever, hitting 50 million downloads within weeks—a velocity that signals aggressive adoption. The model competes directly with Meta's LLaMA 3 and Llama 3.1 70B, but undercuts them on price with zero-cost input and output tokens. Gemma 4 leverages MTP drafters, a technique that Google claims delivers 3x faster inference, as reported in May 2026. Deployment is accelerating: Ollama now supports local runs, and mlx-vlm v0.6.2 added QAT support for Apple Silicon GPUs. Yet critical details remain undisclosed—parameter count, architecture, and licensing terms are still unknown. The model's reliance on Google's ecosystem and MTP drafters creates a dependency that rivals may not replicate. With zero pricing and explosive downloads, Gemma 4 is a momentum play, but the lack of transparency raises questions about long-term viability. Can Google maintain this pace without revealing its technical foundation?
- ·50M downloads in weeks—Google's fastest launch, undercutting Llama with $0 inference cost.
- ·Competes with LLaMA 3 and Llama 3.1 70B; uses MTP drafters for 3x faster inference.
- ·Deployed via Ollama and mlx-vlm for local Apple Silicon support.
- ·Critical specs (parameter count, architecture, licensing) remain undisclosed.
Signal Radar
Five-axis snapshot of this entity's footprint
Mentions × Lab Attention
Weekly mentions (solid) and average article relevance (dotted)
Timeline
6- Research MilestoneApr 30, 2026
Gemma 4 hits 50 million downloads within weeks, fastest Google open model launch
View source - Product LaunchApr 15, 2026
Was integrated by a developer to replace an entire dash cam video analysis stack.
- Product LaunchApr 5, 2026
Community developer ported Gemma 4 to MLX-Swift, enabling local inference on Apple Silicon via LocallyAI app.
View source - Research MilestoneApr 3, 2026
Gemma 4 model demonstrated self-terminating loop detection during a coding task, an emergent behavior for execution control.
View source - Research MilestoneApr 3, 2026
Independent analysis declares Gemma4 models as best-in-class for small open LLMs.
View source- assessment:
- Superior model behavior
Relationships
2Developed
Frequently appears with
4Entities that show up in the same articles — shared coverage, not a stated relationship.
Recent Articles
No articles found for this entity.
Predictions
1- incorrectmonthApr 6, 2026
Google will ship a Gemini 3.x on-device/consumer-hardware release within 2 weeks
Gemma 4 is now surging and the live web context shows Google positioning it explicitly for phones, consumer GPUs, and agentic workflows. The graph cascade from Gemma 4 to Gemini 3.1 and Gemini 3 Deep Think suggests Google is using Gemma as the open-model proving ground before a Gemini-branded follow-on release.
58%
AI Discoveries
1- observationactive2d ago
Lifecycle: Gemma 4
Gemma 4 is in 'declining' phase (0 mentions/3d, 0/14d, 20 total)
90% confidence