Google releases Gemma 4, its most intelligent open-weight model family, available for self-hosting and fine-tuning.
Google announced Gemma 4, the latest generation of its open-weight model series, positioning it as the most capable Gemma release to date. The models are available for download and deployment, continuing Google's strategy of releasing open-weight alternatives to its proprietary Gemini lineup. Gemma 4 competes directly with Meta's Llama series and Mistral in the open-weight space. Specific parameter counts and benchmark scores were not detailed in the announcement but the release follows the established Gemma pattern of multiple size variants.
Gemma 4 is a deployable, self-hostable model you can run on your own infra today — no API costs, no rate limits, no data leaving your stack. If your current pipeline uses Llama or Mistral, this is a direct benchmark challenger worth testing against your eval suite immediately. Open weights also mean fine-tuning is on the table without licensing friction.
Pull the Gemma 4 model via Hugging Face this week and run it against your existing benchmark prompts — measure latency and output quality against your current open-weight baseline to decide if a swap is justified.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Run: pip install transformers accelerate and then open a Python script
Paste: from transformers import AutoTokenizer, AutoModelForCausalLM; model_id = 'google/gemma-4-it'; tokenizer = AutoTokenizer.from_pretrained(model_id); model = AutoModelForCausalLM.from_pretrained(model_id, device_map='auto'); inputs = tokenizer('Explain transformer attention in one paragraph', return_tensors='pt'); outputs = model.generate(**inputs, max_new_tokens=200); print(tokenizer.decode(outputs[0]))
Compare the output quality and generation speed against your current model baseline
Tags