The Open-Source AI Stack
RSS
All models

Models · gemma

Gemma 4 31B

Open weights Google DeepMind · 2026-04-02 · Apache-2.0

Flagship of the Gemma 4 open-weights family, released April 2 2026 under Apache 2.0 alongside a 26B MoE (activates 3.8B per token), an E4B effective-4B edge model, and an E2B. Built off Gemini 3 with native vision and audio, 140-plus language support, and up to 256K context. Google reported the 31B as the #3 open model on Arena AI's text leaderboard at launch.

Cost

$0.00 / Mtok input
$0.00 / Mtok output

· as of 2026-05-21

source ↗

Speed

35.5 tok/sec output
1061 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed × N layers Attention (not disclosed) Position encoding not disclosed context 262,144 tokens Dense MLP SwiGLU activation (standard) 31B active params Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
dense
Total params
31B
Active params
31B
Context window
262K tokens
Attention
unknown
Position encoding
unknown
Post-training
sft, rlhf
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

GPQA-Diamond 85.7 as of 2026-05-21 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
GPTQ Post-training 4-bit weight quantization for GPU serving. runs on vLLM, SGLang, Transformers
EXL2 ExLlamaV2's variable-bitrate format for consumer GPUs. runs on ExLlamaV2
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM
bitsandbytes On-the-fly NF4 / INT8 weight quantization inside Transformers. runs on Transformers

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Apache 2.0 across the full Gemma 4 size ladder
  • · Native audio input on edge variants
  • · 256K context on the larger sizes
  • · Built off the Gemini 3 line

Lineage

31B dense flagship of the Gemma 4 family, built off the Gemini 3 line; sibling to a 26B MoE.

Sources