Gemma
hosomaki.gemma
Open-weight text generation models from Google DeepMind, built from the research behind Gemini and released in sizes that run on a single machine. Executed here by the Ollama runtime.
A Docker Compose file for Gemma alone, generated from the catalog and checked by the safety inspector. Sign in and use Kitchen Credits to activate it.
Uses the Medusa cart and payment-session path. This test control does not create an order or capture payment.
Gemma is a family of open-weight language models published in several parameter sizes, with base and instruction-tuned variants alongside task-specific derivatives such as CodeGemma and PaliGemma. The weights are downloaded and executed locally rather than called over a hosted API, so prompts and completions stay on the deploying machine. In this catalog the executing runtime is Ollama, which handles quantisation, local model storage and the HTTP interface. A first deployment pulls one Gemma variant and issues a completion against it; larger variants need proportionally more memory and are impractical without a GPU.
You know it worked when
- The runtime lists the Gemma model as present in its local library.
- A completion request returns generated text rather than a model-not-found error.
- Token generation proceeds at a usable rate for the chosen variant on the deployment's hardware.
- A second, smaller variant can be pulled and queried without disturbing the first.
Known sharp edges
- The first run downloads several gigabytes of weights and nothing responds until the pull completes.
- Memory demand scales with parameter count and quantisation level, so a variant that loads at 4-bit may fail to load at 8-bit on the same host.
- CPU-only generation works but is slow enough that interactive use is impractical for the larger variants.
- The Gemma Terms of Use restrict certain applications and those restrictions attach to derivatives, which matters before any fine-tune is redistributed.