Gemma 4 31B IT

A July 13, 2026 Google Cloud Community article by Vipul Raja provides a deep technical look at Gemma 4 31B IT, explicitly naming the instruction-tuned 31B variant released by Google DeepMind in April 2026 as part of the Gemma 4 family. The piece breaks down the naming convention (Gemma family, 4th generation, 31 billio The same article details key capabilities directly tied to the 31B IT variant: a configurable "thinking" mode for step-by-step problem-solving on complex tasks, native support for function calling and structured output suited to agentic workflows, and an Apache 2.0 license that allows developers and researchers to free

Google DeepMind released Gemma 4 on April 2, 2026 as a family of multimodal open-weight models supporting text and image input, with audio supported on smaller variants, as detailed in Qubrid's technical deep dive. The family spans four sizes, including the Gemma 4 26B MoE (26B total, 3.8B active) and Gemma 4 31B Dense (31B total and active), both featuring a 256K-token context window and text, image, and video modalities. The post emphasizes the release as a significant open-model launch offering more reasoning, true multimodality, agentic tooling, and a commercially permissive license. The Qubrid write-up situates Gemma 4 within an open-weight ecosystem of more than 400 million downloads and over 100,000 community variants, citing Google DeepMind's framing of "unprecedented intelligence-per-parameter" for advanced reasoning and agentic workflows. Smaller "Effective" variants Gemma 4 E2B and E4B use Dense + PLE architectures with 128K context and full audio/video support, targeting phones and edge devices. The 31B Dense variant targets a single 80GB H100 in bfloat16, while the 26B MoE targets quantized consumer GPUs, with multilingual coverage spanning over 140 languages.

Gemma 4 31B IT is an open-weight, instruction-tuned model from Google released on April 2, 2026, according to Vals AI's benchmark page. The page lists it with a 262k context window, 33k max output tokens, and multimodal input support for text, images, and files. Reported independent benchmark results include 55.03% on SAGE (ranking 2 of 91), 39.33% on Terminal-Bench 2.0, and 61.37% on MortgageTax, with overall accuracy near 51.91% and latency around 5 minutes 30 seconds. Default provider parameters include temperature 1, top-p default, top-k default, 32,768 max output tokens, and high reasoning effort. The Vals AI listing positions Gemma 4 31B IT alongside other open-weight releases and shows rankings restricted to that peer group. Token costs are reported as $0.00 in and $0.00 out, and Google is named as both developer and default provider. Benchmark figures carry explicit error margins such as ±3.41 for SAGE and ±5.21 for Terminal-Bench 2.0, and the page tracks updates including an April 9, 2026 entry following the initial April 2 release date. The model is available for download and comparison across Vals AI's open-weight coverage.

NVIDIA published a NIM model card for Gemma 4 31B IT confirming release on 04/02/2026 via Hugging Face. The card confirms 30.7B total parameters, a 262,144-token vocabulary, and transformer architecture with Proportional RoPE (p-RoPE) and unified Keys and Values in global layers for long-context performance. The model accepts text, image, and video inputs (video processed as frame sequences) and is designed for text generation, reasoning, coding, multimodal understanding, and function calling. NVIDIA explicitly notes the model was developed by Google DeepMind, not NVIDIA, and is available under the Apache 2.0 license.

Google DeepMind released Gemma 4 31B IT as the largest variant in the Gemma 4 open-weights family, licensed under Apache 2.0. It is a dense, instruction-tuned multimodal model handling text and image input and generating text output, with a 256K-token context window and support for over 140 languages. The model ships with configurable thinking modes for reasoning, native function-calling support, and native system prompt handling. It uses a hybrid attention mechanism that interleaves local sliding-window attention with full global attention, targeting deployment on consumer GPUs and workstations for coding, agentic, and multimodal tasks.

A technical deep dive on Gemma 4 describes the family as multimodal, handling text and image input (with audio on small models) and generating text output, with a context window of up to 256K tokens and multilingual support across 140+ languages. The release includes both pre-trained and instruction-tuned variants. The Google DeepMind's stated positioning is "unprecedented intelligence-per-parameter" purpose-built for advanced reasoning and agentic workflows. Since the first Gemma generation, the ecosystem has seen over 400 million downloads and more than 100,000 community variants. The article frames Gemma 4 against DeepSeek R2, Qwe

Google Cloud announced Gemma 4 availability on Vertex AI on April 2, 2026, positioning it as the company's "most capable family of open models" built from the same research as Gemini 3 and released under an Apache 2.0 license. The blog explicitly names the Gemma 4 31B dense model as a variant suited for "complex enterp The post also notes that the Gemma 4 26B MoE variant will become fully managed and serverless on Model Garden "over the coming days," and highlights integration with the Agent Development Kit (ADK) for building and deploying AI agents. Enterprise-focused features include deployment across Sovereign Cloud solutions for