Language modelActive

Gemma 4 31B

Google

Released: March 31, 2026
Data date: September 8, 2026
Predecessor: Gemma 3 27B

Speed: 35.3 tokens/s
Output throughput
Price: $0.15 · $0.4
Input · Output / 1M tokens
Input: Text, Image
Output: Text

Gemma 4 31B is distinguished by its dense architecture, 30.7 billion parameters, and hybrid attention across local and global windows. It is the direct open counterpart to the sparse MoE checkpoint Gemma 4 26B A4B.

Google releases the weights under Apache 2.0. The model card documents text and image input plus reasoning, but no audio. Google defines no hosted Gemini API model ID for the weights.

Specifications and access

Specification	Value and source
Weights	Open weightsSource
Access	API, LocalSource
Parameters	30.7BSource
Context window	262,144 tokensSource
Knowledge cutoff	January 2025Source
Release	2026-03-31Source
Parameters	30.7 BSource
Input	Text, ImageSource
Output	TextSource
Architecture	Dense transformer, 30.7B total parameters; hybrid sliding/global attentionSource
License	Apache 2.0Source
Knowledge cutoff	2025-01Source
Languages	140+Source
Tool calling	YesSource
Announcement	2026-04-02Source

Published benchmarks

MortgageTax

Overall

1.Claude Opus 572.06%

2.Claude Fable 5.170.79%

3.Claude Opus 4.770.27%

4.Claude Sonnet 570.03%

5.Claude Opus 4.869.91%

6.Gemini 3.1 Pro Preview69.4%

7.Claude Fable 568.92%

8.Gemini 2.5 Pro68.92%

9.GPT-5.568.76%

10.Gemini 3 Flash Preview68.72%

11.Claude Opus 4.568.68%

11.Gemini 3.5 Flash-Lite68.68%

13.MiniMax M368.36%

14.GPT-5.468.32%

15.Qwen 3.6 27B68.28%

16.Gemini 3.5 Flash68.12%

17.Gemini 3.1 Flash-Lite68.04%

18.Claude Sonnet 4.667.73%

19.Gemini 3.6 Flash67.61%

20.GPT-5.6 Terra67.33%

21.GPT-5.6 Luna67.29%

22.GPT-5.6 Sol67.29%

23.GPT-5 mini66.89%

24.Gemini 3.7 Flash66.65%

25.Kimi K2.566.53%

26.Kimi K366.34%

27.Qwen 3.7 Plus66.18%

28.GPT-4.165.94%

29.Kimi K2.665.82%

30.o365.7%

31.GPT-4.1 mini65.5%

32.GPT-565.45%

33.Gemini 3.8 Flash65.34%

34.Qwen 3.8 27B64.94%

35.o4-mini64.83%

36.Grok 4.664.19%

37.Inkling63.99%

37.Qwen 3.8 Max 090263.99%

39.GPT-5.4 mini63.51%

40.Inkling Small62.28%

41.Grok 4.561.8%

42.Gemma 4 31B61.37%

43.Mistral Small 3.161.21%

44.MiMo-V2.559.26%

45.GPT-5.4 nano59.1%

46.Llama 4 Scout57.75%

47.GPT-4o mini54.49%

48.GPT-5 nano53.62%

49.GPT-4.1 nano52.82%

50.Mistral Large 352.11%

51.Grok 4.348.25%

52.Grok 4.20 Reasoning45.35%

53.Mistral Medium 3.528.9%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-mortgage-tax:1:overall
Metric: accuracy
Scale: 0-100
Participants: 53
Methodology and context: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-01. Latency in seconds: 41.928. Cost per task in USD: 0. Source ID: vals-ai. Evidence capture ID: vals-mortgage_tax-2026-09-29-8e235fe66b37. Source payload hash: 8e235fe66b379887578d1b76970297d22fc5ae7232f4abfa613cb8576995ae2e. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

SAGE

Overall

1.Claude Opus 4.756.1%

2.Gemma 4 31B55.03%

3.Claude Opus 4.854.79%

4.Kimi K354.26%

5.GPT-5.6 Sol52.56%

6.Qwen 3.8 27B52.4%

7.Claude Fable 551.89%

8.Gemini 3 Flash Preview51.85%

9.Claude Sonnet 5.551.77%

10.GPT-5.551.53%

11.Qwen 3.8 Max 090251.25%

12.GPT-5.4 mini50.81%

13.MiniMax M350.57%

14.Kimi K2.650.22%

15.Gemini 3.5 Flash49.89%

16.Kimi K2.549.87%

17.Gemini 3.1 Flash-Lite49.54%

18.Claude Opus 549.43%

19.Gemini 3.7 Flash49.23%

20.Claude Sonnet 548.93%

21.Gemini 3.6 Flash48.69%

22.Gemini 3.1 Pro Preview48.68%

23.Claude Fable 5.148.53%

24.GPT-6 Luna48.09%

25.DeepSeek-V4.1-Flash47.88%

26.Gemini 3.5 Flash-Lite47.3%

27.GPT-5.6 Terra47%

28.Claude Sonnet 4.646.58%

29.GPT-6 Astra46.37%

30.Claude Opus 5.545.83%

31.Qwen 3.6 27B45.58%

32.Claude Opus 4.545%

33.GPT-6 Sol44.79%

34.GPT-5.6 Luna44.22%

35.GPT-543.68%

36.GPT-5.443.31%

37.MiMo-V2.543.27%

38.GPT-5 mini42.99%

39.Gemini 2.5 Pro41.92%

40.o341.77%

41.o4-mini41.06%

42.Grok 4.740.79%

43.Qwen 3.7 Plus39.25%

44.Grok 4.20 Reasoning38.24%

45.GPT-5.4 nano38.08%

46.Mistral Medium 3.537.61%

47.Inkling36.56%

48.Gemini 3.8 Flash35.06%

49.Grok 4.534.97%

50.Llama 4 Scout34.83%

51.Claude Sonnet 4.532.88%

52.Inkling Small32.74%

53.GPT-5 nano30.38%

54.Grok 4.628.9%

55.Gemini 2.5 Flash-Lite27.53%

56.Mistral Large 324.6%

57.Grok 4.319.74%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-sage:1:overall
Metric: accuracy
Scale: 0-100
Participants: 57
Methodology and context: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 93.058. Cost per task in USD: 0. Source ID: vals-ai. Evidence capture ID: vals-sage-2026-09-29-89f30a3444b7. Source payload hash: 89f30a3444b7edab27c895a52e2ce3759c99ee458f9763715a7a344cb2456193. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Terminal-Bench 2.0

Overall

1.GPT-5.573.2%

2.Claude Opus 4.870.04%

3.Claude Opus 4.768.54%

4.Gemini 3.1 Pro Preview67.42%

4.Gemini 3.5 Flash67.42%

6.GPT-5.3-Codex64.05%

7.Claude Sonnet 4.659.55%

8.Qwen 3.7 Max59.18%

9.Claude Opus 4.558.43%

9.GPT-5.458.43%

11.Kimi K2.657.3%

12.DeepSeek-V4-Pro56.18%

13.GLM-5.153.93%

14.Gemini 3 Flash Preview51.69%

15.GLM-549.44%

16.MiniMax M2.747.19%

17.MiniMax M346.07%

18.GPT-5.4 mini44.94%

18.Qwen 3.6 27B44.94%

20.Grok 4.343.45%

21.MiniMax M2.541.57%

22.Grok 4.20 Reasoning40.45%

22.Kimi K2.540.45%

24.GPT-5.4 nano39.89%

25.Gemma 4 31B39.33%

26.GLM-4.738.2%

27.GPT-537.08%

27.Kimi K2 Thinking37.08%

27.MiniMax M2.137.08%

30.DeepSeek-V3.234.83%

31.Gemini 2.5 Pro30.34%

31.Mistral Medium 3.530.34%

33.GLM-4.628.09%

34.GPT-5 mini26.97%

35.Kimi K2 071125.84%

36.Gemini 3.1 Flash-Lite24.72%

37.DeepSeek-V3.122.47%

38.GPT OSS 120B19.1%

39.Trinity Large Thinking17.98%

40.Mistral Small 416.85%

41.GPT-4.114.61%

42.Mistral Large 38.99%

43.Command A2.25%

higher is betterAs of September 29, 2026Scale: 0% - 100%

Source: Vals AI

CC BY 4.0

gradually.ai

Methodology and context

Comparison cohort: vals-terminal-bench-2:2.0:overall
Metric: accuracy
Scale: 0-100
Participants: 43
Methodology and context: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-06-04. Latency in seconds: 855.298. Source ID: vals-ai. Evidence capture ID: vals-terminal-bench-2-2026-09-29-020c4135b825. Source payload hash: 020c4135b82502d8cc0324b9b7d9474024a86da8f4770c78743bd13103cdfa0c. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source: Vals AI

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

CaseLaw v2: Overall (Show measurement, test conditions, and source)

Source value: 52.626
Score: 52.626
Metric: accuracy
Unit: %
Benchmark version: 2
Category: legal
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-case-law-v2:2:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-05-04. Latency in seconds: 28.716. Cost per task in USD: 0. Source ID: vals-ai. Evidence capture ID: vals-case_law_v2-2026-09-29-cd282b756eb4. Source payload hash: cd282b756eb41a617c01694c5ed3b3d1872a9f417d8aac675370f4943c6c7c63. Parser version: vals-astro-v2. Imported at: 2026-09-29
Context: The source dataset marks this benchmark as archived. The result remains visible as historical evidence and is not interchangeable with an active leaderboard.

Source Vals AI

MMLU Pro (Show measurement, test conditions, and source)

Source value: 85.2
Score: 85.2
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

AIME 2026 (no tools) (Show measurement, test conditions, and source)

Source value: 89.2
Score: 89.2
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

LiveCodeBench v6 (Show measurement, test conditions, and source)

Source value: 80
Score: 80
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

GPQA (Show measurement, test conditions, and source)

Source value: 84.3
Score: 84.3
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

Tau2 (Show measurement, test conditions, and source)

Source value: 76.9
Score: 76.9
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

Humanity's Last Exam (no tools) (Show measurement, test conditions, and source)

Source value: 19.5
Score: 19.5
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

Humanity's Last Exam (search) (Show measurement, test conditions, and source)

Source value: 26.5
Score: 26.5
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

BigBench Extra Hard (Show measurement, test conditions, and source)

Source value: 74.4
Score: 74.4
Metric: percent
Unit: %
Category: vendor-evaluation
Direction: Higher is better
Source type: vendor-reported
Evaluator: Google DeepMind
Status: active
Retrieved at: 2026-09-08
Methodology: Vendor-reported result. Gemma 4 model-card benchmark table; 31B column

Source Google DeepMind

Recorded measurement details

MortgageTax: Overall (Show measurement, test conditions, and source)

Source value: 61.368
Score: 61.368
Metric: accuracy
Unit: %
Benchmark version: 1
Category: finance
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-mortgage-tax:1:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-01. Latency in seconds: 41.928. Cost per task in USD: 0. Source ID: vals-ai. Evidence capture ID: vals-mortgage_tax-2026-09-29-8e235fe66b37. Source payload hash: 8e235fe66b379887578d1b76970297d22fc5ae7232f4abfa613cb8576995ae2e. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

SAGE: Overall (Show measurement, test conditions, and source)

Source value: 55.034
Score: 55.034
Metric: accuracy
Unit: %
Benchmark version: 1
Category: education
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-sage:1:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: private. Leaderboard date: 2026-09-26. Latency in seconds: 93.058. Cost per task in USD: 0. Source ID: vals-ai. Evidence capture ID: vals-sage-2026-09-29-89f30a3444b7. Source payload hash: 89f30a3444b7edab27c895a52e2ce3759c99ee458f9763715a7a344cb2456193. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Terminal-Bench 2.0: Overall (Show measurement, test conditions, and source)

Source value: 39.326
Score: 39.326
Metric: accuracy
Unit: %
Benchmark version: 2.0
Category: agentic
Task: overall
Task label: Overall
Direction: Higher is better
Comparison cohort: vals-terminal-bench-2:2.0:overall
Source type: independent-evaluator
Evaluator: Vals AI
Status: active
Retrieved at: 2026-09-29
Methodology: Source model ID: google/gemma-4-31b-it. Source model name: Gemma 4 31B. Evaluator: Vals AI. Dataset type: public. Leaderboard date: 2026-06-04. Latency in seconds: 855.298. Source ID: vals-ai. Evidence capture ID: vals-terminal-bench-2-2026-09-29-020c4135b825. Source payload hash: 020c4135b82502d8cc0324b9b7d9474024a86da8f4770c78743bd13103cdfa0c. Parser version: vals-astro-v2. Imported at: 2026-09-29

Source Vals AI

Page 1 of 3

Predecessor and successor

Sources and data date

Type	Evidence and data date
Speed	Google. Output speed after the first streaming chunk, not total response time. Retrieved September 8, 2026. Source: Artificial Analysis · Artificial Analysis (retrieved September 8, 2026)Speed measurement methodology
Price	Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. Source: Pricing · API pricing
Input	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Knowledge cutoff data (retrieved September 8, 2026) · Gemma 4 model card · Editorial description reviewed October 2, 2026
Output	Documented modalities. Retrieved September 8, 2026. Source: Model documentation · Knowledge cutoff data (retrieved September 8, 2026) · Gemma 4 model card · Editorial description reviewed October 2, 2026
Research status	Research date September 8, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	active-parameters-billions Original wording Active parameter count is not separately published for this dense model.
Not found in the inspected sources	api-model-id Original wording Gemma is distributed as open weights; the card does not define a hosted Gemini API model ID.
Not found in the inspected sources	training-tokens Original wording The Gemma 4 model card describes the dataset but does not publish a token-count total.
Additional source	Model metadata · Gemma 4 31B instruction-tuned weights · Editorial description reviewed October 2, 2026
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Vals AI (retrieved September 29, 2026)
Additional source	Model metadata (retrieved September 8, 2026)
Additional source	Model metadata (retrieved September 8, 2026)
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Creative Commons Attribution 4.0

Gemma 4 31B

Specifications and access

Published benchmarks

MortgageTax

SAGE

Terminal-Bench 2.0

Measurements without matching peer values

Recorded measurement details

Our tests

Pelican on a bicycle

Popularity

Published prices

Predecessor and successor

Sources and data date