model release

Google releases Gemma 4 family under Apache 2.0 license with 2B to 31B models

TL;DR

Google has released Gemma 4, a family of four open models ranging from 2B to 31B parameters, now available under the Apache 2.0 license for the first time. The 31B dense model ranks 3rd on the Arena AI Text Leaderboard, while the 26B mixture-of-experts variant ranks 6th, both outperforming significantly larger competitors. All models support multimodal inputs and are available on Hugging Face, Kaggle, and Ollama.

2 min read
0

Gemma 4 31B at a glance

Context window262K tokens
Input$0.2/1M tokens
Output$0.4/1M tokens

How Gemma 4 31B ranks now

BenchmarkScoreRankSource
Arena Elo1452 Elo58 of 146LMArena, Nov 2026 ↗
GPQA85.7%79 of 165Artificial Analysis, Jul 2026 ↗
Hallucination (Summarization) (lower is better)7.4%14 of 49Vectara, Nov 2026 ↗
MMLU-Pro85.2%25 of 76llm-stats, Oct 2026 ↗
LiveCodeBench80.0%20 of 48Hugging Face model card ↗

Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.

All Gemma 4 31B specs and scores →

Google has released Gemma 4, its most capable open model family, marking a significant licensing shift: all four models now ship under the commercially permissive Apache 2.0 license, replacing the restrictive Google proprietary license used for earlier Gemma versions.

Model Lineup and Specifications

The Gemma 4 family consists of four variants:

  • E2B and E4B: Effective 2B and 4B parameter models optimized for edge devices (smartphones, Raspberry Pi, Jetson Orin Nano). Both support 128K token context windows and natively handle images, video, and audio input.
  • 26B MoE (Mixture-of-Experts): 3.8 billion active parameters with up to 256K token context. Designed for latency-optimized inference.
  • 31B Dense: Maximum quality variant with up to 256K token context, intended as a foundation model for fine-tuning.

All models are multimodal, with the larger variants supporting vision inputs. The architecture is based on the same technology powering Google's proprietary Gemini 3.

Performance and Benchmarks

The 31B model currently ranks 3rd among all open models on the Arena AI Text Leaderboard with a score above 1,440 Elo, while the 26B MoE ranks 6th. According to Artificial Analysis, on the GPQA Diamond benchmark for scientific reasoning:

  • 31B model: 85.7% in reasoning mode—second-best among open models under 40B parameters (behind Qwen3.5 27B at 85.8%)
  • 26B MoE model: 79.2%—ahead of OpenAI's gpt-oss-120B (76.2%)

Google claims the 31B model outperforms models 20 times its size. The evaluation requires approximately 1.2 million output tokens, less compute than comparable competitors.

Hardware Deployment and Framework Support

The unquantized bfloat16 weights of the 31B model fit on a single 80GB NVIDIA H100 GPU, with quantized versions suitable for consumer graphics cards. The 26B MoE model activates only 3.8 billion parameters during inference for efficient token generation.

Gemma 4 supports deployment across:

  • Hardware: NVIDIA (Jetson Orin Nano through Blackwell), AMD (ROCm), and Google TPUs (Trillium, Ironwood)
  • Frameworks: Hugging Face Transformers, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM/NeMo, LM Studio, Unsloth, SGLang, and Keras
  • Platforms: Hugging Face, Kaggle, Ollama, Google AI Studio (31B/26B), and Google AI Edge Gallery (E4B/E2B)

Production Capabilities

All models natively support function calling, structured JSON output, and system instructions for agentic workflows. Fine-tuning is available through Google Colab, Vertex AI, or local GPUs. Production deployments scale via Google Cloud (Vertex AI, Cloud Run, GKE).

What This Means

The Apache 2.0 licensing removes commercial restrictions that previously limited Gemma adoption, making these models viable for proprietary applications and enterprise use without derivative work constraints. The 31B model's 3rd-place ranking on Arena AI—achieved with one-tenth the parameters of many competitors—demonstrates that parameter efficiency, not scale, now determines practical performance. For edge computing, the E2B/E4B variants with multimodal support directly compete with mobile-optimized models from Mistral and others. The shift signals Google's commitment to competing in open models while maintaining architectural parity with Gemini 3.

Related Articles

model release

Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context

Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.

model release

Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license

Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

model release

Google releases Gemini 4 Neon at $3/$15 per 1M tokens with a 1M-token context window

Google released Gemini 4 Neon, available today through the API at $3 per million input tokens and $15 per million output tokens. It is the fourth new Gemini 4 model in four weeks, with a context window of about one million tokens.

Comments

Loading...