Ollama Gemma Check

Status: tested Date: 2026-05-20

Result

Ollama is installed and running on this machine.

Tested models:

gemma3:270m
gemma4:e4b

Gemma through Ollama works for plain chat through this runtime:

python3 -m agentic_loop --provider ollama --model gemma3:270m \
  "Reply with exactly: OK"

Result:

OK

The adapter command output also includes a note when tool schemas had to be disabled for this model.

But gemma3:270m does not support native Ollama tool calling. Ollama returned:

registry.ollama.ai/library/gemma3:270m does not support tools

The adapter now detects that and falls back to a no-tool model call instead of crashing.

On 2026-05-21, gemma4:e4b was tested with native tool calling:

python3 -m agentic_loop chat \
  --provider ollama \
  --model gemma4:e4b \
  --enable-network-tools

The prompt well, now tell me the news today top news returned Google News RSS results through search_news.

ollama ps reported:

gemma4:e4b    11 GB    100% GPU    32768 context

That confirms GPU offload for the loaded model. It does not prove that every GPU device is active; use host GPU tooling such as nvidia-smi for per-device utilization when the driver is visible to the shell.

The runtime also has a repeated-tool-call guard: if a tool-capable model repeats the exact same successful tool call, the controller finalizes from the previous tool result instead of running until the max-step limit.

Meaning

Gemma can be used as a local chat model in this runtime.

Gemma 3 270M cannot currently be used as the native tool-calling planner for agentic actions in Ollama. For native Ollama tool calling, use a tool-capable model such as gemma4:e4b or another Ollama model that advertises tool support.

Check Command

python3 scripts/smoke_ollama.py --model gemma3:270m

Observed smoke result:

{
  "model": "gemma3:270m",
  "native_tool_calling_for_inspect": false,
  "ollama_chat_works": true
}