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

Gemma 4 31B IT

Available on OpenRouterOn OpenRouterAvailable nowGenerally availableCan run on your own serversOpen weightsOur pick: The best starting point for training on your own textsPick: Best open model to fine-tune

Gemma 4 31B 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. If you want a model that writes in your organisation's voice, Gemma 4 31B is the best place to start: Google publishes the untrained base version, good training guides exist, and it already handles Swedish well.

Open-weights dense 30.7B. Configurable thinking mode (on/off). No Google list API price.

Writing & creativity: —
Research & analysis: —
Coding: —
Price: Cheap$0.09 / $0.34
Price per 1M (blended)Blended / 1M: $0.15
MemoryContext: 262K
Longest answerMax output: —
Test resultsResults: 24 (11 independent11 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-4-31b-it.

Route it as google/gemma-4-31b-it at $0.09 in / $0.34 out per 1M tokens, 262K context. Listed since 2 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): 30.7B
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 ≈ 32 GB at 8-bit, 17 GB at 4-bit (+10–30% for KV cache).

Languages: Languages: Card: out-of-the-box support for 35+ languages, pretrained on 140+ (Swedish not named). EuroEval Swedish: IT 1.58, base 2.27.

Training it further: Fine-tuning: Google docs: QLoRA fine-tuning with Hugging Face Transformers/TRL (ai.google.dev/gemma/docs/core/huggingface_text_finetune_qlora); Unsloth 'Fine-tune Gemma 4' guide. Pretrained base published.

Quantisations: bf16, qat q4_0 gguf (google/gemma-4-31B-it-qat-q4_0-gguf), qat w4a16 (google/gemma-4-31B-it-qat-w4a16-ct), nvfp4 (NVIDIA), gguf/mlx (community: unsloth, lmstudio). Engines: Transformers, llama.cpp.

Download (Hugging Face)Weights on Hugging Face ↗huggingface.co ↗ai.google.dev ↗ai.google.dev ↗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
AA-LCR	69.7%	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
AA-Omniscience Index	-47.9	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b; AA-Omniscience Index (-100..100)
AIME 2026	89.2%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	No tools.
Artificial Analysis Intelligence Index	14.7	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
Artificial Analysis output speed	36 tok/s	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b; median output tokens/sec across AA-tracked API providers (not a self-hosted measurement)
BIG-Bench Extra Hard	74.4%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
Codeforces Elo	2150	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Elo.
EuroEval Swedish (generative)	1.6	Defaultgoogle/gemma-4-31B-it (val)	Independent testIndependentEuroEval (Alexandra Institute) ↗	29 Sep 2026	—	EuroEval rank tier 4; ±0.14; lower is better; task scores (first metric): SweDN summarisation 39.58 ± 0.15, Skolprov 58.00 ± 4.64, Swedish facts 36.59 ± 2.52, ScaLA-sv 71.36 ± 1.15
GPQA Diamond	85.7%	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
GPQA Diamond	84.3%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
Humanity's Last Exam	23.6%	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
Humanity's Last Exam	19.5%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	No tools.
Humanity's Last Exam (with tools)	26.5%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	With search.
LiveCodeBench	80.0%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	LiveCodeBench v6.
LLM Creative Story-Writing Benchmark (Lech Mazur)	-1.9	DefaultGemma 4 31B Reasoning	Independent testIndependentLech Mazur (LLM Creative Story-Writing Benchmark) ↗	28 Sep 2026	—	rank 44/56; Thurstone comparison score (centered at 0); est. win chance 24%; 95% bootstrap -1.985 to -1.778
LMArena Text (overall)	1452	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Arena AI (text) as of 2026-04-02.
MMLU-Pro	85.2%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
MMMLU	88.4%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
MMMU-Pro (no tools)	76.9%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—
OpenAI MRCR v2 (8-needle)	66.4%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	8 needle, 128k average.
SciCode	45.5%	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
tau2-bench	76.9%	DefaultThinking	Maker's own figureVendor-reportedGoogle ↗	2 Apr 2026	—	Average over 3 domains.
Terminal-Bench 2.1	43.5%	DefaultGemma 4 31B (Reasoning)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug gemma-4-31b
Vectara Hallucination Leaderboard (HHEM)	7.4%	Defaultgoogle/gemma-4-31b-it	Independent testIndependentVectara ↗	22 Sep 2026	—	factual consistency 92.6 %; answer rate 100.0 %; avg summary 75.8 words; HHEM-2.3 judge; effort not stated (API default)