Gemma 4 12B
- Input
- $0.10/M tok
- Output
- $0.30/M tok
- Blended
- $0.15/M tok · 3:1
- Speed
- 118tok/s
- Context
- 262Ktokens
Long-tail import from Artificial Analysis (2026-09-24); AA variant: Gemma 4 12B (Non-reasoning). Price = AA median host price.
Rankings
Writing & chat
Verified#97 / 162rank by quality35.7±9.5quality / 1001/5 benchmarks measuredPrice frontierVision
Verified#72 / 98rank by quality35.7±8.2quality / 1001/6 benchmarks measuredHard reasoning
Verified#152 / 218rank by quality29.2±5.9quality / 1002/4 benchmarks measuredAgents
Verified#113 / 187rank by quality24.3±10.0quality / 1001/7 benchmarks measured
Self-hostingEstimated
How we estimate →What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.
- Hardware
- 1× B200 180GB
- BF16 weights
- Throughput
- ~8,514 tok/s
- many requests batched
- Price
- $0.10–0.91 /M tok
- busy → light use
- On your own machine
- RTX 5090 32GB (int4, ~210 tok/s single-stream)
- single consumer GPU or Mac
Assumptions (5)
- BF16 weights (11.95B params) + 25% KV-cache headroom ≈ 30 GB
- 1× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
- ~8,514 output tok/s aggregate at batch 243 (bandwidth-bound)
- Blended 3:1 input:output; prefill ~44,482 tok/s
- Low = 75% utilization at the low GPU price; high = 20% at the high price
Benchmark scores
~ italic, dashed = no published score yet, estimated from related benchmarks.
Writing & chat
MMMLUEstimated
~last of 5 (estimated)below measured range (<81.3%)
No published score — estimated from related benchmarks
LMArena Text, Non-English (Elo)Estimated
~#44 of 57 (estimated)~1362 Elo
- #1Gemini 4 ArgonFew votes1512
- #2Claude Opus 5.51497
- #43Qwen3 Next 80B A3B Instruct1386
- ~#44Gemma 4 12BEstimated1362
- #44GLM-4.6VFew votes1361
No published score — estimated from related benchmarks
LMArena Text (Elo)Estimated
~#44 of 57 (estimated)~1381 Elo
- #1Gemini 4 Argon1525
- #2Claude Opus 5.51507
- #43Qwen3 Next 80B A3B Instruct1399
- ~#44Gemma 4 12BEstimated1381
- #44GLM-4.6VFew votes1378
No published score — estimated from related benchmarks
EQ-Bench Creative Writing v3 (Elo)Estimated
~#22 of 26 (estimated)~1461 Elo
- #1GPT-6 Astra2173
- #2Claude Fable 5.12162
- #21Gemini 3.1 Pro1491
- ~#22Gemma 4 12BEstimated1461
- #22Mistral Large 31412
No published score — estimated from related benchmarks
Vision
CharXiv ReasoningEstimated
~last of 12 (estimated)below measured range (<74.5%)
No published score — estimated from related benchmarks
OmniDocBench v1.5Estimated
~#6 of 8 (estimated)~78.2%
- #1MiniMax-M391.6%
- #2Kimi K391.1%
- #5Claude Haiku 4.579.6%
- ~#6Gemma 4 12BEstimated78.2%
- #6Mistral Small 476.4%
No published score — estimated from related benchmarks
GDP.pdf (AA)Estimated
~#38 of 48 (estimated)~5.4%
- #1Muse Wonder Full32.6%
- #2GPT-6 Astra31.0%
- #37Qwen3.6 35B A3B6.6%
- ~#38Gemma 4 12BEstimated5.4%
- #38Claude Haiku 4.53.8%
No published score — estimated from related benchmarks
ChartographyEstimated
~last of 25 (estimated)below measured range (<9.0%)
No published score — estimated from related benchmarks
Hard reasoning
GPQA DiamondVerified
#124 of 19266.1%
- #1GPT-6 Astra96.1%
- #2Gemini 3.8 Flash95.3%
- #124Falcon-H1R-7B66.1%
- #124Gemma 4 12B66.1%
- #126Solar Open 100B65.7%
artificialanalysis.ai · 2026-09-24
Humanity's Last ExamVerified
#141 of 1986.3%
- #1Claude Fable 5.274.3%
- #2Claude Opus 5.561.4%
- #139Magistral Small 1.26.4%
- #141Gemma 4 12B6.3%
- #141Olmo 3.1 32B Think6.3%
artificialanalysis.ai · 2026-09-24
AIME (latest)Estimated
~#23 of 34 (estimated)~56.9%
- #1GLM-5.3 NVFP499.5%
- #2Beam 1.0 501B A23B97.7%
- #22Phi-4-reasoning-plus78.0%
- ~#23Gemma 4 12BEstimated56.9%
- #23Mistral Large 338.0%
No published score — estimated from related benchmarks
AA Intelligence IndexEstimated
~#85 of 103 (estimated)~11.1
- #1Gemini 4 Carbon59.4
- #2Claude Opus 5.557.6
- #83Ling 3.0 Tiny11.1
- ~#85Gemma 4 12BEstimated11.1
- #85Trinity Large Thinking10.8
No published score — estimated from related benchmarks
Agents
τ²-bench (Telecom)Verified
#59 of 11931.9%
- #1Step 3.7 Flash98.5%
- #2DeepSeek-V4-Pro-0813-nvfp4-DSpark98.3%
- #58Sarvam 30B34.5%
- #59Gemma 4 12B31.9%
- #60GLM-4.6V30.7%
artificialanalysis.ai · 2026-09-24
OSWorld-VerifiedEstimated
~#12 of 15 (estimated)~55.8%
- #1Kimi K384.8%
- #2Qwen3.8 27B84.3%
- #11Muse Glimmer65.9%
- ~#12Gemma 4 12BEstimated55.8%
- #12Nemotron 3.5 Super VL NVFP452.3%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~#3 of 5 (estimated)~53.8%
No published score — estimated from related benchmarks
BrowseCompEstimated
~#20 of 22 (estimated)~44.5%
- #1Atria Dawn Preview92.5%
- #2GPT-6 Astra91.5%
- #19Mistral Medium 3.548.6%
- ~#20Gemma 4 12BEstimated44.5%
- #20Nemotron 3 Ultra44.4%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Estimated
~#57 of 92 (estimated)~747 Elo
- #1Claude Opus 5.51866
- #2Muse Wonder Full1857
- #56Mistral Medium 3.5747
- ~#57Gemma 4 12BEstimated747
- #57Ring-2.6-1T734
No published score — estimated from related benchmarks
AutomationBench-AAEstimated
~#49 of 85 (estimated)~6.5%
- #1Claude Fable 5.276.4%
- #2Muse Wonder Full71.3%
- #48Muse Glimmer6.8%
- ~#49Gemma 4 12BEstimated6.5%
- #49Mistral Medium 3.56.3%
No published score — estimated from related benchmarks
τ-Bench Banking (AA)Estimated
~#62 of 91 (estimated)~10.7%
- #1Muse Spark 1.350.5%
- #2GLM-5.350.3%
- #61Devstral Small 210.7%
- ~#62Gemma 4 12BEstimated10.7%
- #62Devstral 210.5%
No published score — estimated from related benchmarks