Gemma 4 model variants — picker guide¶

Which Gemma 4 variant should you load? Depends on hardware, use-case, and whether you care more about latency or response quality. This doc tabulates the trade-offs and gives you the single-line answer for the common cases.

TL;DR¶

You're...	Pick
Evaluating Duecare on a laptop	`gemma4:e2b` (default)
NGO Mac mini with 16 GB RAM	`gemma4:e2b`
NGO Mac mini with 32 GB RAM, ≥ 5 caseworkers	`gemma4:e4b`
Cloud Run / Render free tier	`gemma4:e2b-int4` (fastest cold start)
GPU pod (T4 / L4 / A10) for production	`gemma4:e4b`
Old phone (4 GB RAM)	`gemma4:e2b-int4`
Modern Android (8 GB+ RAM)	`gemma4:e2b` (the v0.9 default)
Workstation with H100 / A100	`gemma4:31b`

All Gemma 4 variants we support¶

Variant	Params	Quant	On-disk	Min RAM	RPS / vCPU	Latency p95 (E2B chat ≈ 200 tok)	License
`gemma4:e2b`	2 B	INT8	1.5 GB	8 GB	0.5	4-8 s (CPU) / 0.5-1 s (T4)	Apache 2.0
`gemma4:e2b-int4`	2 B	INT4	750 MB	4 GB	0.7	3-6 s (CPU) / 0.3-0.7 s (T4)	Apache 2.0
`gemma4:e4b`	4 B	INT8	3.5 GB	16 GB	0.25	8-16 s (CPU) / 1-2 s (T4)	Apache 2.0
`gemma4:e4b-int4`	4 B	INT4	2.0 GB	8 GB	0.4	6-12 s (CPU) / 0.7-1.5 s (T4)	Apache 2.0
`gemma4:31b`	31 B	Q4_0	18 GB	32 GB	n/a (GPU only)	1-3 s (A100)	Apache 2.0

Numbers measured with tests/load/k6_chat.js against Ollama on the same host. Real RPS depends on your prompt + max-tokens + hardware; treat ±50%.

Compatibility map¶

Where it runs	Variants supported
Ollama (CPU)	all of the above
Ollama (NVIDIA GPU)	all of the above
Ollama (Apple Silicon Metal)	all of the above
MediaPipe LiteRT-LM (Android)	`gemma4:e2b-int4`, `gemma4:e2b` (INT8), `gemma4:e4b-int4`, `gemma4:e4b` (INT8)
llama.cpp / GGUF	`gemma4:e2b`, `gemma4:e4b`, `gemma4:31b` (Q4_0/Q8_0/F16)
HuggingFace Transformers	all (some need PEFT / transformers HEAD)
HuggingFace Inference Endpoint	`gemma4:e2b`, `gemma4:e4b` (paid endpoints)

The Android app's ModelManager ships all four MediaPipe Gemma 4 variants selectable from Settings, each with multiple mirror-fallback URLs.

When Gemma 4 features become load-bearing¶

Gemma 4 has three features the harness leans on:

Native function calling. The Coordinator agent (packages/duecare-llm-agents/src/duecare/agents/coordinator/) uses Gemma 4's function-calling protocol to dispatch among the GREP / RAG / Tools / Persona layers. Earlier Gemma generations couldn't reliably emit valid JSON tool-call payloads at this scale; v0.6+ assumes Gemma 4. On an earlier-generation fallback, the coordinator drops to a regex-based "tool intent" parser with measurably lower precision.
Multimodal input. The Scout agent reads contract / receipt photos via Gemma 4's image-encoder path. Gemma 2 doesn't do this at all; PaliGemma 2 does but adds a second model load. v0.6+ keeps Scout as a single-model path on Gemma 4.
Native long context. Gemma 4's 128k token context window lets the journey-aware prompt include the worker's full journal (median 5-15k tokens) without truncation. Gemma 2's 8k window forced ad-hoc summarization that lost evidence.

For these reasons the default everywhere is now Gemma 4 E2B — earlier Gemma generations are kept only as documented fallbacks and lose these features.

Picking a variant in practice¶

As an NGO director with a Mac mini¶

# Mac mini M2 8 GB — Gemma 4 E2B fits with headroom
DUECARE_OLLAMA_MODEL=gemma4:e2b docker compose up -d

# Mac mini M2 32 GB or M2 Pro — go E4B for noticeably better answers
DUECARE_OLLAMA_MODEL=gemma4:e4b docker compose up -d

As an enterprise platform on cloud GPUs¶

# Single-tenant T4 / L4 — E4B is sweet-spot
helm install duecare ./infra/helm/duecare \
  --set chat.env.DUECARE_OLLAMA_MODEL=gemma4:e4b \
  --set chat.nodeSelector."cloud\.google\.com/gke-accelerator"=nvidia-tesla-t4

# Multi-tenant A100 pool — 31B for premium tier, E4B for standard
# (configure per-tenant routing via feature_flags.py)

As an Android app worker¶

The app's Settings → On-device model lets the worker pick from six variants. Default is E2B INT8; the app's mirror-fallback list tries litert-community/gemma-4-E2B-it-litert-lm then litert-community/gemma-4-E2B-it then a github.com/TaylorAmarelTech/duecare-journey-android/releases/download/models-v1/ mirror.

Switching at runtime¶

The default docker-compose.yml reads DUECARE_OLLAMA_MODEL from .env. To switch:

# Edit .env or set inline
DUECARE_OLLAMA_MODEL=gemma4:e4b docker compose up -d ollama
docker compose exec ollama ollama pull gemma4:e4b
docker compose restart chat

The chat tier sees the new model on the next request — no rebuild.

Future variants¶

Gemma 4 Multilingual (Q3 2026 expected) — drop-in replacement for E4B with extended language coverage. Will be added to the picker once published.
Gemma 4 Instruct LoRA/DPO candidates — internal Duecare adapters should now be produced through the A-00 preconfigured pipeline and recorded in the exported report bundle before any HF Hub publication. Not yet shipping.

When new variants land, this doc updates + kaggle/_INDEX.md + the Android app's ModelManager.kt get the new entries together.