Gemma 4 Uncensored

• - 🆕 Uncensored fine-tune built on Google's Gemma 4 26B MoE base
• - 🧠 Mixture-of-Experts: 26B total, 4B active per token
• - 📏 256K-token context window for long conversations and documents
• - 👁️ Supports vision input, function-calling and web-search tooling
• - 🔧 Ships in int4 quantization for efficient local deployment
• - 🔒 Refusal-removal fine-tune for filter-free chat
• - 💬 Targets coding and general-purpose conversation
• - 🏢 Derived from Google's Gemma 4 family, catalog-dated April 2026

Google

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Gemma 4 Uncensored is a community-style derivative of Google's Google Gemma 4 26B A4B Instruct, the Mixture-of-Experts member of the Gemma 4 lineup. Per the catalog description, the 26B MoE variant activates only about 4 billion of its parameters per token, an arrangement designed to deliver fast throughput while retaining the footprint of a larger model. This build ships in int4 quantization and exposes a 256K-token context window.

Relative to earlier uncensored builds derived from Google Gemma 3 27B Instruct, the Gemma 4 foundation reflects several concrete capabilities in this catalog entry: vision input, native function-calling and web-search tooling, alongside the Mixture-of-Experts architecture and the expanded context window. The "uncensored" label reflects fine-tuning or abliteration that removes the base model's refusal behavior, targeting filter-free chat, coding and general-purpose conversation. Independent uncensored Gemma 4 26B fine-tunes are published on Hugging Face.

Google offers a companion dense variant, Google Gemma 4 31B Instruct, the larger member of the same generation. Within this catalog, a later sibling, Gemma 4 26B A4B Uncensored, appears as a subsequent variant in the same uncensored lineage. As with any refusal-removed model, long-context behavior and edge cases may be incompletely characterized.