Gemma 4 E4B Coding Edition — Release & Publishing Notes

What this release is

gemma4-e4b-code is a coding-oriented Ollama derivative built on the user's existing local gemma4:e4b model.

The Modelfile keeps the base weights unchanged and adds:

• A coding-focused system instruction.
• 32K default context for a practical 16 GB unified-memory baseline.
• 16,384 maximum generated tokens.
• Gemma 4's recommended sampling defaults: temperature 1.0, top_p 0.95, top_k 64.
• min_p 0.05 for modest output filtering.
• Thinking enabled through Gemma 4's <|think|> system control token.

Gemma 4 E4B itself supports a 128K context window. The 32K Modelfile default is deliberate: context size has a direct memory cost, so a portable public model should not force 128K on every computer.

Download from ollama:

ollama pull mukunthpr/gemma4-e4b-code

Then skip to the adaptive runtime section

Build from Gemma 4 E4B

ollama pull gemma4:e4b
ollama create gemma4-e4b-code -f Modelfile
ollama run gemma4-e4b-code

Verify the model:

ollama show gemma4-e4b-code
ollama ps

Adaptive macOS runtime

The included run-gemma4-code-mac.sh is a runtime launcher for the model to tune it to your system specifications

Run it locally with:

./run-gemma4-code-mac.sh

Or point the same launcher at the published model:

./run-gemma4-code-mac.sh YOUR_USERNAME/gemma4-e4b-code

The script detects installed physical memory, starts a temporary local Ollama server with the selected runtime settings, and launches the same model through that server. The model name remains unchanged across machines.

Context profiles:

Installed RAM	Selected context
<16 GiB	16K
16–23 GiB	32K
24–31 GiB	48K
32–47 GiB	64K
48–63 GiB	96K
64+ GiB	128K

These are conservative operating profiles, not hard hardware limits. Heavy applications, large projects, multimodal inputs, and multiple simultaneously loaded models can require lowering the context size.

Optional overrides:

OLLAMA_CODE_CTX=65536 ./run-gemma4-code-mac.sh YOUR_USERNAME/gemma4-e4b-code
OLLAMA_CODE_MAX_OUTPUT=8192 ./run-gemma4-code-mac.sh YOUR_USERNAME/gemma4-e4b-code

The runner uses a dedicated local port so that the adaptive settings apply to the runtime it starts without changing the user's normal Ollama server configuration.

Ollama memory tuning

The runner enables:

OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q8_0
OLLAMA_NUM_PARALLEL=1

Flash Attention can significantly reduce memory use as context grows. Ollama documents q8_0 KV cache as using approximately half the memory of f16 with a small precision impact. OLLAMA_NUM_PARALLEL=1 prevents concurrent requests from multiplying the effective context allocation.

---

Made with ❤️ by Mukunth P.R