Gemma 4 12B

Open weights256k context

Gemma 4 12B is an open-weights model from Google with a 256k-token context window. No provider we track lists it yet.

Providers serving Gemma 4 12B

No provider we track lists this model yet.

Gemma 4 12B benchmark scores

Compare every model in Bench rankings

Published scores only, each linked to its source. Ticks are every other model's best published score.

Each score is the model's best published result across its settings, named beside it. A setting above the default can use more tokens per task, which a per-token price does not show.

Show as table
Benchmark scores for Gemma 4 12B, with the setting each was measured at
BenchmarkScoreSetting
GPQA Diamond
PhD-level scientific reasoning, pass@1, no tools
75.3%, source: Artificial Analysis, thinking onthinking on
MMMU-Pro (standard)
Visual reasoning, 10 answer options
69.7%, source: Artificial Analysis, thinking onthinking on
Humanity's Last Exam (text-only)
Reasoning & knowledge, no tools
15.7%, source: Artificial Analysis, thinking onthinking on
τ²-Bench Telecom
Agentic tool use (telecom), with a Qwen3 model as the simulated user
36.3%, source: Artificial Analysis, thinking onthinking on
Terminal-Bench 2.1 (Terminus 2)
Agentic coding & terminal use
27.3%, source: Artificial Analysis, thinking onthinking on
IFBench
Instruction following, 294 prompts, 5 runs each
73.5%, source: Artificial Analysis, thinking onthinking on
GDPval-AA v2.1 (Elo)
Real-world work tasks; a rating against other models, not a percentage
427, source: Artificial Analysis, thinking onthinking on
Terminal-Bench Hard
Agentic coding & terminal use, Artificial Analysis's 44-task hard subset
18.2%, source: Artificial Analysis, thinking onthinking on
LCR
Long-context reasoning, AA-LCR v1.1
63.7%, source: Artificial Analysis, thinking onthinking on
Omni Accuracy
Omniscience knowledge accuracy
15.6%, source: Artificial Analysis, thinking onthinking on
Non-Hallucination
1 − hallucination rate
26.6%, source: Artificial Analysis, thinking offthinking off

More from Google

See all 16 Google models

QAI lists these prices; it does not serve or resell these models. Prices are each provider's standard on-demand list price in US dollars per million tokens (default region, standard tier, peak hours, base context tier), checked on the date shown against the provider's own published prices or, where it has no public price page, its listing on OpenRouter. Most benchmark scores come from Artificial Analysis; Epoch AI (GPQA Diamond), MathArena (AIME 2026) and the APEX-Agents and CritPt leaderboards add the rest, and where none of them has a model, or none ran it at full capability, a figure its maker published may stand in (marked v); in some benchmarks it was run the maker's own way, which its note names. Each benchmark follows one test setup, the one used by the source that covers the most models, and leaves results from other setups out. Each score is the best its source published across the model's settings, names any non-default setting, and links to its source.

An AIME 2026 score noted “released after the exam” comes from a model released after the exam, so the problems may have been in its training data. GDPval-AA is a rating against other models (Elo), not a percentage. Data as of 26 Oct 2026.

Sources and licences