Overview
Run Gemma 4 31B IT on Crusoe. Try Google’s 31B open model with image input, 256K-token context, and Apache 2.0 licensing, ready for fine-tuning and inference.
Explore additional models
Frequently
asked questions
Gemma 4 31B IT is good for instruction following, multilingual chat, summarization, and assistant use cases that benefit from Google’s safety tuning, and it accepts both text and image input for document and visual understanding. It is a dense instruction-tuned model with 32.7B parameters and a 262,144-token context. On Crusoe it is available for Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning. You can also request a Tailored Deployment for bespoke inference optimization and a dedicated endpoint.
Gemma 4 31B IT is a dense model that uses all of its parameters on every token, while Gemma 4 26B is a mixture-of-experts model that activates about 4B parameters per token. The 31B ranks higher on quality, placing third among open models on LMArena’s text leaderboard versus sixth for the 26B, while the 26B is faster. Crusoe offers both, so you can run the higher-quality 31B IT through serverless access or on a dedicated GPU.
Gemma 4 31B IT is billed by how you use it. Serverless Inference is billed per token, with separate rates for uncached input, cached input, and output tokens, and no minimum commitment. Self-Serve Deployments are billed per GPU hour, and Serverless Fine-Tuning per 1M training tokens. Access it through an OpenAI-compatible API in Crusoe Intelligence Foundry. Current rates are shown in the pricing module above.
The “IT” in Gemma 4 31B IT stands for instruction-tuned. Google releases each Gemma model in two versions: a pretrained (“pt”) base model that predicts the next token, and an instruction-tuned (“it”) model further trained to follow instructions, hold conversations, and apply Google’s safety tuning. The instruction-tuned version is the right choice for chat, assistants, and most production applications.
Yes, Gemma 4 31B IT is available for Serverless Fine-Tuning in Crusoe Intelligence Foundry. Run a LoRA job on a JSONL or Parquet dataset, then deploy the checkpoint in one click to Self-Serve Deployments or download the weights.
MemoryAlloy technology is a proprietary memory fabric in Crusoe’s inference engine that shares the KV cache across nodes, with cache-aware routing that sends each request to the node most likely to have its context ready. For Gemma 4 31B IT, repeated system prompts, shared documents, and multi-turn conversations skip redundant prefill, so the model handles more concurrent traffic at consistent latency, and reused context is billed at the lower cached-input rate.


%204.png)
