Google · 4B · Local

Gemma 4 e4B tools · the rule-breaker

Every other 4B model is chat-only — tool calling normally needs 7–8B. Google's "efficient" e4B breaks that rule: it uses Aperio's memory, files and web in under 3 GB of RAM. Not bulletproof, but genuinely capable for its size.

Size: 4B (efficient)
Comfortable RAM: 8 GB
Memory used: ~2.5–3 GB
Tool calling: ~75% — works, retry sometimes
Best for: Tools on minimal hardware; a light second model
Avoid if: You need 3+ tool chains or rock-solid reliability

New here? The two-minute concepts primer explains parameters, tool calling and the rest — they apply to every model, so they aren't repeated here.

Hardware

The lightest tool-capable model there is. It runs comfortably where a full 8B model won't fit, which is the whole reason to choose it.

Apple Silicon · 8 GB+Even a base M1 MacBook Air runs it at full speed. On 16–32 GB machines it's barely a blip — light enough to run alongside a bigger model.
PC · NVIDIA GPU, 4 GB+ VRAMA GTX 1050 Ti 4 GB handles it; an RTX 2060 is more than enough.
PC · CPU onlySnappy even without a GPU — roughly 15–25 words/sec on a recent i5 or Ryzen 5.

Give it room for conversation in .env:

LLAMACPP_CTX=32768

Try it yourself

These tasks prove the tools work — and show you the retry habit that makes a 4B tool-user practical. When a call doesn't fire, naming the tool usually fixes it.

1 · The moment of truth — memory across chats

This is the rule-break. A 4B model storing a fact and recalling it in a brand-new conversation was impossible a generation ago.

Paste thisPlease remember this: my favourite season is autumn, and I grew up in a town called Hirakata.

Then, new chatOpen a fresh conversation and paste: "What do you know about me from memory?"

✅ GoodIt confirms saving the memory, then in the new chat recalls autumn + Hirakata. You're watching a 4B model call remember and recall.

❌ The wallIt says "I can't store memories," or recalls nothing. At 4B this happens maybe 1 in 4 times.

→ If it failsBe explicit: "Use the recall tool to look up any memories about me." Naming the tool fixes most misses at this size.

2 · Read a file

Single-tool file access — its steadiest trick once memory works.

Paste thisRead the README.md file in the project directory and summarise it for me in two sentences.

✅ GoodIt calls read_file, then gives a two-sentence summary of the actual contents.

❌ The wallIt invents a summary without reading, or asks you to paste the file. Retry naming read_file.

→ MeansIf this works, it can read your notes and documents — one file at a time.

3 · A two-step tool chain

Two tools in one prompt — the practical ceiling at 4B. More than two and it starts dropping steps.

Paste thisRemember that I'm working on a project called "Balcony Herbs." Then read the README.md in the project directory and tell me one thing in it that's relevant to a beginner.

✅ GoodIt stores the project, reads the file, and answers — both tools, right order.

❌ The wallIt does one and forgets the other, or muddles the order. Normal at the 4B frontier.

→ MeansKeep chains to two steps. For longer workflows, step up to Llama 3.1 8B.

4 · Reach past its training cutoff

Its knowledge is frozen (~mid-2025), but unlike the chat-only 4B models it can fetch the live web.

Paste thisSearch the web and tell me who won the most recent FIFA World Cup, with the year and the source.

✅ GoodIt uses fetch_url, returns the current answer and a source.

❌ The wallIt answers from stale memory without fetching. Retry: "Use the web fetch tool to look it up."

→ MeansWith web fetch working, the cutoff stops mattering for facts you can look up.

The verdict

Gemma 4 e4B is impressive engineering: real tool calling at a fraction of the usual memory cost. Choose it when you want memory and file access but can't fit a full 8B model, or want a light second model for background jobs. Just budget for the occasional retry — at 4B, tool calling works most of the time, not every time.

Where to go next

More reliable? → Llama 3.1 8B~85% tool reliability and 2–3 tool chains, at 16 GB. Just chatting? → Small modelsIf you don't need tools, the chat-only 4B models are lighter still. Compare everything →The full lineup in one table.