Model Drop

Gemma 4

April 2, 2026 2 min read

Gemma 4, which Google DeepMind released on April 2, 2026, is a family of four open-weight multimodal reasoning models, led by a 31B dense model and a 26B mixture-of-experts, with E2B and E4B models for phones and edge devices.

Model: Gemma 4 31B (google/gemma-4-31B-it), plus 26B A4B, E4B and E2B
Maker: Google DeepMind
Ship date: April 2, 2026
Pricing: Free (open weights)
Context: 256K tokens (E2B and E4B: 128K)
Weights: Open (Apache 2.0)
Where you can use it:
- Hugging Face, Kaggle and Ollama weights
- Google AI Studio (31B and 26B A4B)
When to reach for it:
- Local reasoning and coding on one GPU
- E2B and E4B for phones and edge devices
- Skip 31B and 26B for audio input
How it compares:
- Open rival to Qwen3.5 27B and gpt-oss-120b
- Trails GLM-5 and Kimi K2.5 in chat votes
What makes it different:
- Gemma now ships under plain Apache 2.0
- Image and video input; audio on E2B/E4B
Early reactions: Simon Willison got "probably the best pelican" he'd seen from a laptop model out of the 26B A4B, though the 31B GGUF was broken in LM Studio on day one.
Read more: Gemma 4: Byte for byte, the most capable open models · Gemma 4 31B model card

The benchmarks

Google's numbers are for the instruction-tuned models with thinking on, compared only with its own Gemma 3 27B (no thinking) and the Gemma 4 siblings, so Gemma 4 31B leads every row Google printed. The launch blog's table labels the 85.2% row MMMLU, but the model card lists 85.2% as MMLU Pro (and 88.4% as MMMLU), so the model card's label is used. The τ²-bench Retail row and the Arena score as of April 2, 2026 come from the blog table; the rest are from the launch model card. The Arena scores are Google's own; LMArena's published April 2 data has the Gemma 4 models a few points lower.

Google DeepMind's numbers

MMLU-Pro (harder general knowledge): 85.2%. +2.6 over Gemma 4 26B A4B (82.6%), +17.6 over Gemma 3 27B (67.6%).
GPQA Diamond (grad-level science, no tools): 84.3%. +2.0 over Gemma 4 26B A4B (82.3%), +41.9 over Gemma 3 27B (42.4%).
AIME 2026 (math contest, no tools): 89.2%. +0.9 over Gemma 4 26B A4B (88.3%), +68.4 over Gemma 3 27B (20.8%).
LiveCodeBench (coding problems, v6): 80.0%. +2.9 over Gemma 4 26B A4B (77.1%), +50.9 over Gemma 3 27B (29.1%).
τ²-bench Retail (agent tool use): 86.4%. +0.9 over Gemma 4 26B A4B (85.5%), +79.8 over Gemma 3 27B (6.6%).
Humanity's Last Exam (no tools) (expert questions): 19.5%. +10.8 over Gemma 4 26B A4B (8.7%).
MMMU Pro (multimodal reasoning): 76.9%. +3.1 over Gemma 4 26B A4B (73.8%), +27.2 over Gemma 3 27B (49.7%).
Arena AI text score (Google-reported chat votes): 1452. +11 over Gemma 4 26B A4B (1441), +87 over Gemma 3 27B (1365).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 14.7. −2.0 under Gemma 4 26B A4B (16.7), −8.2 under Qwen3.5 27B (22.9), +3.1 over gpt-oss-120b (11.6).