upcube
Google•Open weights•June 3, 2026

Gemma 4 12B

Intelligence, Performance & Price Analysis

Source:Artificial Analysis scrapeType:Open weights
Intelligence
22

score

Gemma 4 12B (Reasoning) scores 22 (estimated) on the Artificial Analysis Intelligence Index, …

vs class avg: 9

Speed
121.9

output tokens/sec

Gemma 4 12B (Reasoning) generates output at 121.9 tokens per second (based on the median across …

Latency
2.37s

TTFT

Gemma 4 12B (Reasoning) has a time to first token (TTFT) of 2. …

Input Price#78/130
$0.10

/ 1M tokens

Gemma 4 12B (Reasoning) costs $0.10 per 1M input tokens (better than average, median: $0.18) and $0. …

Output Price#82/130
$0.30

/ 1M tokens

Gemma 4 12B (Reasoning) costs $0.10 per 1M input tokens (better than average, median: $0.18) and $0. …

Decision Overview

Ethen's intelligence-driven routing assessment for this model

Overall verdict

Strong general-purpose model

Suitable for most production tasks, but high-volume or repetitive work should still be compared against cheaper routes.

Best for

  • Code generation
  • Content analysis
  • Structured extraction
  • Customer-facing chat
  • Low-latency chat

Weak or unsuitable

  • Extremely long-context tasks
  • Ultra-low-latency requirements

Routing recommendation

1

Use the fit matrix to compare against cheaper routes when volume rises

Cost pressure

Competitive pricing — $0.10 / M input, $0.30 / M output. Cost pressure is low enough for sustained production use.

Model-fit matrix

Use this model where capability pays for itself; route away where volume, latency, or budget dominate.

Use caseFitReasonRouting note
Complex reasoning & agentic workflowsStrongIntelligence score 22 supports capable reasoning, but very hard tasks may benefit from higher-tier models.Good for most complex tasks; consider a frontier model for the hardest 10%.
High-volume chat & customer-facingExcellentOutput speed 121.9 tokens/sec and capable intelligence make this suitable for real-time chat at scale.Ideal for interactive chat; enable caching for repeated queries.
Latency-sensitive applicationsStrongTTFT 2.37s — adequate latency for most interactive use cases.Suitable for real-time; test with your specific workload.
Cost-sensitive pipelinesExcellentOutput pricing at $0.30 is very competitive for high-volume workloads.Excellent for budget-constrained pipelines; enable caching to reduce costs further.

Cost pressure

Low

Pricing is competitive — input $0.10, output $0.30. Suitable for sustained production use.

Cheaper substitutes

No cheaper substitute is supported by the current price data.

Route away when

  • Extreme scale where even low costs matter

Highlights

Key benchmark results at a glance.

Intelligence

Intelligence position

Open section

Artificial Analysis Intelligence Index · Higher is better · Evaluation results measured independently by Artificial Analysis

Gemma 4 12B (Reasoning) scores 22 (estimated) on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 9).

0204060Claude Fable 5 (with fallback): 6060AClaude Fable 5 (with fal…GPT-5.5 (xhigh): 5555AIGPT-5.5 (xhigh)GLM-5.2 (max): 5151?GLM-5.2 (max)Gemini 3.5 Flash: 5050GGemini 3.5 FlashMiniMax-M3: 4444?MiniMax-M3DeepSeek V4 Pro (max): 4444DDeepSeek V4 Pro (max)Kimi K2.6: 4444KKimi K2.6Muse Spark: 4343?Muse SparkNemotron 3 Ultra: 3838NNemotron 3 UltraGrok 4.3 (high): 3838xGrok 4.3 (high)gpt-oss-120b (high): 2424AIgpt-oss-120b (high)Gemma 4 12B: 2222?Gemma 4 12B

Speed

Speed position

Open section

Output tokens per second · Higher is better · Evaluation results measured independently by Artificial Analysis

Gemma 4 12B (Reasoning) generates output at 121.9 tokens per second (based on the median across providers serving the model), which is above average compared to other open weight models of similar size (median: 101.8 t/s).

088175263350gpt-oss-120b (high): 316316AIgpt-oss-120b (high)Nemotron 3 Ultra: 248248NNemotron 3 UltraGLM-5.2 (max): 216216?GLM-5.2 (max)gpt-oss-20b (high): 197197AIgpt-oss-20b (high)Gemini 3.5 Flash: 192192GGemini 3.5 FlashMistral Small 3.1: 171171MiMistral Small 3.1Qwen3.6 35B A3B: 166166QQwen3.6 35B A3BGrok 4.3 (high): 164164xGrok 4.3 (high)Qwen3 30B A3B 2507: 133133QQwen3 30B A3B 2507Gemma 4 12B: 122122?Gemma 4 12BNVIDIA Nemotron 3 Nano: 116116NNVIDIA Nemotron 3 NanoMiniMax-M3: 9595?MiniMax-M3

Context, Cost & Pricing

Price/value position

Open section

Weighted average cost (USD) per Intelligence Index task · Lower is better · Evaluation results measured independently by Artificial Analysis

Pricing is competitive — input $0.10, output $0.30. Suitable for sustained production use.

$0.00$1.00$2.00$3.00$4.00$5.00DeepSeek V4 Pro (max): $0.04$0.04DDeepSeek V4 Pro (max)gpt-oss-120b (high): $0.06$0.06AIgpt-oss-120b (high)MiniMax-M3: $0.12$0.12?MiniMax-M3Grok 4.3 (high): $0.14$0.14xGrok 4.3 (high)Nemotron 3 Ultra: $0.24$0.24NNemotron 3 UltraKimi K2.6: $0.35$0.35KKimi K2.6GLM-5.2 (max): $0.37$0.37?GLM-5.2 (max)Gemini 3.5 Flash: $0.59$0.59GGemini 3.5 FlashGPT-5.5 (xhigh): $0.86$0.86AIGPT-5.5 (xhigh)Claude Fable 5 (with fallback): $2.75$2.75AClaude Fable 5 (with fal…

Intelligence & Quality

Benchmark scores and quality indices measuring model capability.

AA-Omniscience Index

AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct. · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

-100-50050Claude Fable 5 (with fallback): 4040AClaude Fable 5 (with fal…Gemini 3.5 Flash: 2323GGemini 3.5 FlashGPT-5.5 (xhigh): 2020AIGPT-5.5 (xhigh)Grok 4.3 (high): 1818xGrok 4.3 (high)Kimi K2.6: 6.46.4KKimi K2.6Muse Spark: 4.14.1?Muse SparkGLM-5.2 (max): 44?GLM-5.2 (max)MiniMax-M3: 1.41.4?MiniMax-M3Nemotron 3 Ultra: -0.8-0.8NNemotron 3 UltraDeepSeek V4 Pro (max): -10-10DDeepSeek V4 Pro (max)Gemma 4 12B: -51.9-51.9?Gemma 4 12B

Artificial Analysis Intelligence Index by Open Weights / Proprietary

Artificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0204060Claude Fable 5 (with fallback): 6060AClaude Fable 5 (with fal…GPT-5.5 (xhigh): 5555AIGPT-5.5 (xhigh)GLM-5.2 (max): 5151?GLM-5.2 (max)Gemini 3.5 Flash: 5050GGemini 3.5 FlashMiniMax-M3: 4444?MiniMax-M3DeepSeek V4 Pro (max): 4444DDeepSeek V4 Pro (max)Kimi K2.6: 4444KKimi K2.6Muse Spark: 4343?Muse SparkNemotron 3 Ultra: 3838NNemotron 3 UltraGrok 4.3 (high): 3838xGrok 4.3 (high)Gemma 4 12B: 2222?Gemma 4 12B

Artificial Analysis Intelligence Index

Artificial Analysis Intelligence Index v4.1 incorporates 9 evaluations: GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0204060Claude Fable 5 (with fallback): 6060AClaude Fable 5 (with fal…GPT-5.5 (xhigh): 5555AIGPT-5.5 (xhigh)GLM-5.2 (max): 5151?GLM-5.2 (max)Gemini 3.5 Flash: 5050GGemini 3.5 FlashMiniMax-M3: 4444?MiniMax-M3DeepSeek V4 Pro (max): 4444DDeepSeek V4 Pro (max)Kimi K2.6: 4444KKimi K2.6Muse Spark: 4343?Muse SparkNemotron 3 Ultra: 3838NNemotron 3 UltraGrok 4.3 (high): 3838xGrok 4.3 (high)Gemma 4 12B: 2222?Gemma 4 12B

Artificial Analysis Openness Index: Score

Openness Index assesses model openness on a 0 to 100 normalized scale (higher is more open) · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

050100Nemotron 3 Ultra: 8383NNemotron 3 UltraNVIDIA Nemotron 3 Nano: 8383NNVIDIA Nemotron 3 NanoDeepSeek V4 Pro (max): 5050DDeepSeek V4 Pro (max)Magistral Small 1.2: 5050?Magistral Small 1.2GLM-5.2 (max): 4444?GLM-5.2 (max)Qwen3 30B A3B 2507: 4444QQwen3 30B A3B 2507Gemma 4 12B: 3939?Gemma 4 12Bgpt-oss-120b (high): 3939AIgpt-oss-120b (high)Qwen3.6 27B: 3939QQwen3.6 27BQwen3.6 35B A3B: 3939QQwen3.6 35B A3B

Intelligence

Artificial Analysis Intelligence Index · Higher is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0204060Claude Fable 5 (with fallback): 6060AClaude Fable 5 (with fal…GPT-5.5 (xhigh): 5555AIGPT-5.5 (xhigh)GLM-5.2 (max): 5151?GLM-5.2 (max)Gemini 3.5 Flash: 5050GGemini 3.5 FlashMiniMax-M3: 4444?MiniMax-M3DeepSeek V4 Pro (max): 4444DDeepSeek V4 Pro (max)Kimi K2.6: 4444KKimi K2.6Muse Spark: 4343?Muse SparkNemotron 3 Ultra: 3838NNemotron 3 UltraGrok 4.3 (high): 3838xGrok 4.3 (high)Gemma 4 12B: 2222?Gemma 4 12B

Speed & Throughput

Output tokens per second and generation throughput.

Output Speed

Output tokens per second · Higher is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

088175263350gpt-oss-120b (high): 316316AIgpt-oss-120b (high)Nemotron 3 Ultra: 248248NNemotron 3 UltraGLM-5.2 (max): 216216?GLM-5.2 (max)gpt-oss-20b (high): 197197AIgpt-oss-20b (high)Gemini 3.5 Flash: 192192GGemini 3.5 FlashMistral Small 3.1: 171171MiMistral Small 3.1Qwen3.6 35B A3B: 166166QQwen3.6 35B A3BGrok 4.3 (high): 164164xGrok 4.3 (high)Qwen3 30B A3B 2507: 133133QQwen3 30B A3B 2507Gemma 4 12B: 122122?Gemma 4 12B

Speed

Output tokens per second · Higher is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

088175263350gpt-oss-120b (high): 316316AIgpt-oss-120b (high)Nemotron 3 Ultra: 248248NNemotron 3 UltraGLM-5.2 (max): 216216?GLM-5.2 (max)Gemini 3.5 Flash: 192192GGemini 3.5 FlashGrok 4.3 (high): 164164xGrok 4.3 (high)Gemma 4 12B: 122122?Gemma 4 12BMiniMax-M3: 9595?MiniMax-M3GPT-5.5 (xhigh): 8888AIGPT-5.5 (xhigh)Kimi K2.6: 7676KKimi K2.6DeepSeek V4 Pro (max): 7272DDeepSeek V4 Pro (max)

Latency & Response Time

Time to first token and end-to-end response latency.

End-to-End Response Time

Seconds to output 500 tokens, including reasoning model 'thinking' time · Lower is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0s15s30s45s60sGemma 4 31B: 15s15s?Gemma 4 31BQwen3.5 9B: 8.4s8.4sQQwen3.5 9BQwen3.6 27B: 8.4s8.4sQQwen3.6 27BDevstral Small 2: 7.4s7.4s?Devstral Small 2Kimi K2.6: 6.6s6.6sKKimi K2.6Ministral 3 14B: 6.4s6.4s?Ministral 3 14BMagistral Small 1.2: 6.2s6.2s?Magistral Small 1.2Qwen3 32B: 5.4s5.4sQQwen3 32BMiniMax-M3: 5.2s5.2s?MiniMax-M3NVIDIA Nemotron 3 Nano: 4.3s4.3sNNVIDIA Nemotron 3 NanoGemma 4 12B: 4.1s4.1s?Gemma 4 12B

Latency: Time To First Answer Token

Seconds to first answer token received · Accounts for reasoning model 'thinking' time · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0s15s30s45s60sKimi K2.6: 58s58sKKimi K2.6Gemma 4 31B: 50s50s?Gemma 4 31BQwen3.5 9B: 34s34sQQwen3.5 9BQwen3.6 35B A3B: 33s33sQQwen3.6 35B A3BMagistral Small 1.2: 25s25s?Magistral Small 1.2Qwen3 32B: 21s21sQQwen3 32BMiniMax-M3: 21s21s?MiniMax-M3NVIDIA Nemotron 3 Nano: 17s17sNNVIDIA Nemotron 3 NanoGemma 4 12B: 16s16s?Gemma 4 12BQwen3 30B A3B 2507: 15s15sQQwen3 30B A3B 2507

Context, Cost & Pricing

Context window size, cost per task, and token pricing.

Context Window

Context window: tokens limit · Higher is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0250k500k750k1MGemini 3.5 Flash: 1M1MGGemini 3.5 FlashClaude Fable 5 (with fallback): 1M1MAClaude Fable 5 (with fal…GLM-5.2 (max): 1M1M?GLM-5.2 (max)DeepSeek V4 Pro (max): 1M1MDDeepSeek V4 Pro (max)Grok 4.3 (high): 1M1MxGrok 4.3 (high)MiniMax-M3: 1M1M?MiniMax-M3NVIDIA Nemotron 3 Nano: 1M1MNNVIDIA Nemotron 3 NanoGPT-5.5 (xhigh): 922k922kAIGPT-5.5 (xhigh)Nemotron 3 Ultra: 262k262kNNemotron 3 UltraMuse Spark: 262k262k?Muse SparkGemma 4 12B: 256k256k?Gemma 4 12B

Cost per Intelligence Index Task

Weighted average cost (USD) per Artificial Analysis Intelligence Index task, segmented by token type. Lower is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

00.51GPT-5.5 (xhigh): 0.10.1AIGPT-5.5 (xhigh)Gemini 3.5 Flash: 0.10.1GGemini 3.5 FlashQwen3.6 27B: 00QQwen3.6 27BGLM-5.2 (max): 00?GLM-5.2 (max)Qwen3.6 27B: 00QQwen3.6 27BKimi K2.6: 00KKimi K2.6Qwen3.6 35B A3B: 00QQwen3.6 35B A3BNemotron 3 Ultra: 00NNemotron 3 UltraMiniMax-M3: 00?MiniMax-M3Grok 4.3 (high): 00xGrok 4.3 (high)

Cost per Task

Weighted average cost (USD) per Intelligence Index task · Lower is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

$0.00$1.00$2.00$3.00$4.00$5.00Claude Fable 5 (with fallback): $2.75$2.75AClaude Fable 5 (with fal…GPT-5.5 (xhigh): $0.86$0.86AIGPT-5.5 (xhigh)Gemini 3.5 Flash: $0.59$0.59GGemini 3.5 FlashGLM-5.2 (max): $0.37$0.37?GLM-5.2 (max)Kimi K2.6: $0.35$0.35KKimi K2.6Nemotron 3 Ultra: $0.24$0.24NNemotron 3 UltraGrok 4.3 (high): $0.14$0.14xGrok 4.3 (high)MiniMax-M3: $0.12$0.12?MiniMax-M3gpt-oss-120b (high): $0.06$0.06AIgpt-oss-120b (high)DeepSeek V4 Pro (max): $0.04$0.04DDeepSeek V4 Pro (max)

Cost to Run Artificial Analysis Intelligence Index

Cost (USD) to run all evaluations in the Artificial Analysis Intelligence Index · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

050100150200Qwen3.6 35B A3B: 165165QQwen3.6 35B A3BGPT-5.5 (xhigh): 132132AIGPT-5.5 (xhigh)Gemini 3.5 Flash: 114114GGemini 3.5 FlashQwen3.6 27B: 8888QQwen3.6 27BNemotron 3 Ultra: 3434NNemotron 3 UltraGLM-5.2 (max): 3333?GLM-5.2 (max)Qwen3.6 27B: 2828QQwen3.6 27BKimi K2.6: 2525KKimi K2.6MiniMax-M3: 1515?MiniMax-M3Grok 4.3 (high): 1313xGrok 4.3 (high)

Pricing: Cache Hit, Input, and Output

Price (USD per M Tokens) · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0250k500k750k1MQwen3 32B: 9.19.1QQwen3 32BGLM-5.2 (max): 6.16.1?GLM-5.2 (max)Kimi K2.6: 5.15.1KKimi K2.6Qwen3.6 27B: 4.24.2QQwen3.6 27BQwen3.6 27B: 4.24.2QQwen3.6 27BGrok 4.3 (high): 44xGrok 4.3 (high)Nemotron 3 Ultra: 3.63.6NNemotron 3 UltraQwen3 30B A3B 2507: 2.62.6QQwen3 30B A3B 2507Magistral Small 1.2: 22?Magistral Small 1.2Qwen3.6 35B A3B: 1.71.7QQwen3.6 35B A3BGemma 4 12B: 0.40.4?Gemma 4 12Binputoutputcache Hit

Time per Intelligence Index Task

Weighted average decode time (minutes) per task; excludes TTFT and overhead time · Lower is better · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

02468DeepSeek V4 Pro (max): 6.76.7DDeepSeek V4 Pro (max)Qwen3.6 27B: 6.56.5QQwen3.6 27BGemma 4 31B: 5.55.5?Gemma 4 31BClaude Fable 5 (with fallback): 4.84.8AClaude Fable 5 (with fal…MiniMax-M3: 44?MiniMax-M3NVIDIA Nemotron 3 Nano: 3.23.2NNVIDIA Nemotron 3 NanoQwen3.6 27B: 33QQwen3.6 27BGPT-5.5 (xhigh): 33AIGPT-5.5 (xhigh)GLM-5.2 (max): 2.82.8?GLM-5.2 (max)Qwen3.6 35B A3B: 2.72.7QQwen3.6 35B A3B

Architecture & Scale

Model size, parameters, and architectural details.

Model Size: Total and Active Parameters

Comparison between total model parameters and parameters active during inference · Evaluation results measured independently by Artificial Analysis

Source and methodology are summarized below

Chart source and provenance are listed in Methodology & sources below.

0250k500k750k1MDeepSeek V4 Pro (max): 1.6k1.6kDDeepSeek V4 Pro (max)Kimi K2.6: 1k1kKKimi K2.6GLM-5.2 (max): 753753?GLM-5.2 (max)Nemotron 3 Ultra: 550550NNemotron 3 UltraMiniMax-M3: 428428?MiniMax-M3gpt-oss-120b (high): 117117AIgpt-oss-120b (high)Qwen3.6 35B A3B: 3636QQwen3.6 35B A3BQwen3 32B: 3333QQwen3 32BNVIDIA Nemotron 3 Nano: 3232NNVIDIA Nemotron 3 NanoGemma 4 31B: 3131?Gemma 4 31Bpassive Paramsactive Params

Specifications

Technical Specifications

Gemma 4 12B (Reasoning) scores 22 (estimated) on the Artificial Analysis Intelligence Index, placing it well above average among other open weight models of similar size (median: 9). Gemma 4 12B (Reasoning) generates output at 121.9 tokens per second (based on the median across providers serving the model), which is above average compared to other open weight models of similar size (median: 101.8 t/s). Gemma 4 12B (Reasoning) costs $0.10 per 1M input tokens (better than average, median: $0.18) and $0.30 per 1M output tokens (better than average, median: $0.40), based on the median across providers serving the model.

Gemma 4 12B

Model type
Open weights
Reasoning
Yes
Input modalities
Gemma 4 12B (Reasoning) supports text, image, speech, and video input.
Output modalities
Gemma 4 12B (Reasoning) supports text output.
Context window
260k tokens
Open weights / source
Yes, Gemma 4 12B (Reasoning) is open weights. The model weights are publicly available and can be downloaded for self-hosting.
Parameters
Gemma 4 12B (Reasoning) has 12 billion parameters.
License
Gemma 4 12B (Reasoning) is released under the Apache 2.0 license. This license allows commercial use.
API availability
Yes, Gemma 4 12B (Reasoning) is available via API through 1 provider.

Methodology & Provenance

This page is rendered from the normalized profile and page JSON for Gemma 4 12B.

Benchmark values are preserved as normalized; only layout, disclosure ordering, and typography are adjusted for readability.

Frequently Asked Questions