Models · Google (Gemini / DeepMind) · Out sinceReleased 2 Apr 2026

Gemma 4 26B A4B IT

Available on OpenRouterOn OpenRouterAvailable nowGenerally availableCan run on your own serversOpen weights

Gemma 4 26B A4B IT is made by Google (Gemini / DeepMind). We don't have enough test results yet to rank it. It's cheap to use. Its makers have published it, so you can run it on your own servers.

Open-weights MoE, 3.8B active of ~26B. Configurable thinking mode. No Google list API price.

Writing & creativity: —
Research & analysis: —
Coding: —
Price: Cheap$0.08 / $0.26
Price per 1M (blended)Blended / 1M: $0.12
MemoryContext: 262K
Longest answerMax output: —
Test resultsResults: 16 (3 independent3 indep.)
Out sinceReleased: 2 Apr 2026
Made byVendor: Google (Gemini / DeepMind)
UnderstandsInputs: text, image

Scores are out of 100 for each area, compared with every model we track; “#” is its rank. Context is how much text it can read at once (1M is roughly 700,000 words). Blended price mixes the cost of what you send and what it writes back.

Thinking levels

Reasoning settings

No reasoningDefault

Links

Using it through OpenRouter

On OpenRouter

OpenRouter is a service that gives access to many AI models in one place. This model is listed there as google/gemma-4-26b-a4b-it.

Route it as google/gemma-4-26b-a4b-it at $0.08 in / $0.26 out per 1M tokens, 262K context. Listed since 3 Apr 2026.

frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p

Run it yourself

Self-hosting

On your own servers,Licence, sizeunder your own control.and hardware.

Compare all open models →All open-weights models →

LicenceLicence: Apache-2.0
OK for business useCommercial use: Yes
SizeParams (total / active): 25.2B / 3.8B
Can be trained furtherBase model: Yes ↗

Hardware you'd need

Hardware tier & memory

Runs on a powerful laptop or workstation.

≤ 24 GB at 4-bit (one consumer GPU / 32 GB Mac). Weights ≈ 26 GB at 8-bit, 14 GB at 4-bit (+10–30% for KV cache). MoE: 3.8B active per token.

Languages: Languages: 35+ languages out of the box, pretrained on 140+. EuroEval Swedish: IT 1.77, base 2.53.

Training it further: Fine-tuning: Same as Gemma 4 31B (Google QLoRA guide, Unsloth).

Quantisations: bf16, qat q4_0 gguf (google/gemma-4-26B-A4B-it-qat-q4_0-gguf), nvfp4 (NVIDIA), gguf/mlx (community). Engines: Transformers, llama.cpp.

Download (Hugging Face)Weights on Hugging Face ↗huggingface.co ↗unsloth.ai ↗

Every test result

Every result

All the numbers,with where they came from.

Every result we've found for this model, with who measured it and a link to where we read it.

All raw rows (vendor and independent kept separate), with setting label, source, date, harness and notes.

All results for this model
						NotesNotes
AIME 2026	88.3%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	No tools.
Artificial Analysis Intelligence Index	16.7	DefaultGemma 4 26B A4B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-26b-a4b; AA flags this Intelligence Index as ESTIMATED (intelligenceIndexIsEstimated=true, not all component evals run) - treat as provisional
BIG-Bench Extra Hard	64.8%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
Codeforces Elo	1718	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Elo.
EuroEval Swedish (generative)	1.8	Defaultgoogle/gemma-4-26B-A4B-it (val)	Independent testIndependentEuroEval (Alexandra Institute) ↗	29 Sep 2026	—	EuroEval rank tier 6; ±0.10; lower is better; task scores (first metric): SweDN summarisation 39.10 ± 0.16, Skolprov 50.25 ± 2.83, Swedish facts 41.31 ± 2.92, ScaLA-sv 63.52 ± 1.78
GPQA Diamond	82.3%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
Humanity's Last Exam	8.7%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	No tools.
Humanity's Last Exam (with tools)	17.2%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	With search.
LiveCodeBench	77.1%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	LiveCodeBench v6.
LMArena Text (overall)	1441	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Arena AI (text) as of 2026-04-02.
MMLU-Pro	82.6%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
MMMLU	86.3%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
MMMU-Pro (no tools)	73.8%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
OpenAI MRCR v2 (8-needle)	44.1%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	8 needle, 128k average.
tau2-bench	68.2%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Average over 3 domains.
Vectara Hallucination Leaderboard (HHEM)	5.2%	Defaultgoogle/gemma-4-26b-a4b-it	Independent testIndependentVectara ↗	22 Sep 2026	—	factual consistency 94.8 %; answer rate 99.8 %; avg summary 67.1 words; HHEM-2.3 judge; effort not stated (API default)