Gemma-4-12B-IT
google/gemma-4-12b-it
google/gemma-4-12B-it on HuggingFace ↗
Throughput
Decode tok/s after a fixed context fill (PBM) — same agent-style workload as bench v2.
Recipe
- Profile
- google-gemma-4-12b-it-llama
- Engine
- llama.cpp
- Context
- 256k · q8_0 KV
- Served as
- gemma-4-12b-it
Run on your Spark
single node
shell
spark inference up google-gemma-4-12b-it-llama
Why we run it
Best mid-size Gemma 4 for a single Spark: encoder-free unified multimodal (text/image/audio), 256K context, near-26B-MoE quality at ~12B dense. Apache 2.0, not gated. Prefer -it over base.
Bench notes
PBM 4k @ 19.8 tok/s — perfbench-metrics — profile=google-gemma-4-12b-it-llama
Measurement history
Context ladder
Older bench-v2 / golden cells at each benched context window (single measurement).
| Context | KV | Throughput |
|---|---|---|
| @ 32k peak golden | — | 14.1t/s |
Benchmark runs
Recorded inference benchmark sessions for this model's profile.
| Date | Profile | Method | Avg | Session t/s | Range | Fill | Tool |
|---|---|---|---|---|---|---|---|
| 2026-06-28latest | google-gemma-4-12b-it-llama | bench-agent-v2v2.0 | 14.1t/s | — | — | ~50,000 | ok |
| golden 256k/q8_0 @ 14.1 tok/s — fill~50000 — bench-agent-v2 — tool_ok=True | |||||||
| 2026-06-21 | google-gemma-4-12b-it-llama | bench-agent | 21.1t/s | 21.1 · 21.1 · 20.9 | 20.9–21.1 | @ 32k | — |
| agent bench avg 21.1 tok/s over 3 sessions × 3 turns (2304 tok in 109.4s) | |||||||