Gemma 3n

Spec

Attribute	Value
Developer	Google DeepMind
Released	2026-05-12 (early preview)
Announced	2026-05-12
Context window	unknown
Pricing	unknown
License	unknown
Availability	Google AI Studio (early preview); targeting Qualcomm / MediaTek / Samsung devices
Parameters	5B / 8B (raw)
Active Memory	2 GB / 3 GB (via PLE)
Modalities	Text, Image, Audio, Video
Deployment Target	Mobile / on-device + cloud livestream

Release Date

Early preview released 2026-05-12.

Architecture Innovation: Per-Layer Embeddings (PLE)

Key engineering contribution of Gemma 3n: Per-Layer Embeddings (PLE) dramatically reduce RAM footprint compared to traditional embedding tables that load all at once.

• Gemma 3n 5B: 5B parameters → only 2 GB RAM
• Gemma 3n 8B: 8B parameters → only 3 GB RAM
• Reduction: ~3× less RAM vs conventional approach

The 8B variant contains a nested 2B submodel — the model can dynamically switch between 2B and 4B active memory footprint on the fly, enabling adaptive performance/quality tradeoffs without hosting separate models.

Capabilities

• Audio: ASR (transcription) + Speech Translation (speech → translated text)
• Images: Multimodal understanding
• Video: Enhanced video understanding
• Multilingual: Japanese, German, Korean, Spanish, French (WMT24++ ChrF 50.1%)

Performance

• 1.5× faster on mobile vs Gemma 3 4B (prior gen)
• WMT24++: 50.1% (ChrF multilingual benchmark)
• Built with Qualcomm Technologies, MediaTek, Samsung System LSI for device optimization

Benchmarks

Benchmark	Score
WMT24++ (ChrF)	50.1%
Mobile speed vs Gemma 3 4B	1.5× faster
More benchmarks pending full release.

Use Cases

• On-device AI for smartphones (Qualcomm / MediaTek / Samsung)
• Real-time ASR and translation on device
• Cloud livestreaming with low-latency multimodal processing
• Developer experimentation via Google AI Studio

Compared To

Model	Parameters	RAM	Deployment	Notes
Gemma 3n 5B	5B raw	2 GB	Mobile/edge	PLE architecture
Gemma 3n 8B	8B raw	3 GB	Mobile/edge	Nested 2B submodel
Gemma 3 4B	4B	~4 GB+	Edge/cloud	Previous gen
Devstral Small 2	24B	~14 GB	RTX/DGX Spark	Coding focus
Phi-4 (Microsoft)	14B	~8 GB	Edge/PC	General capability

Sources

• Developer guide: deepmind.google/blog/introducing-gemma-3n-the-developer-guide/ (local snapshot)
• Google Developers Blog: developers.googleblog.com/en/introducing-gemma-3n/
• Google DeepMind X: x.com/GoogleDeepMind/status/1925916216083779774

Related

• Google DeepMind
• Gemini 3.1 Deep Think — DeepMind's frontier reasoning model (same org)
• AlphaEvolve — DeepMind algorithm design agent