Gemma-4-26B-A4B-IT
google/gemma-4-26b-a4b-it
google/gemma-4-26B-A4B-it on HuggingFace ↗
Throughput
Decode tok/s after a fixed context fill (PBM) — same agent-style workload as bench v2.
Recipe
- Profile
- google-gemma-4-26b-a4b-it-eugr
- Engine
- vLLM
- Context
- 256k · fp8 KV
- Served as
- gemma-4-26b-a4b-it
Run on your Spark
single node
shell
spark inference up google-gemma-4-26b-a4b-it-eugr
Why we run it
Popular MoE pick (~3.8B active / 26B total) — same efficiency class as Qwen3.6 MoE. 256K context, text+image. Heavier than 12B but fast inference per token.
Bench notes
PBM 4k @ 23.0 tok/s — perfbench-metrics — profile=google-gemma-4-26b-a4b-it-eugr
Measurement history
Context ladder
Older bench-v2 / golden cells at each benched context window (single measurement).
| Context | KV | Throughput |
|---|---|---|
| @ 8k peak golden | — | 22.2t/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-4-26b-a4b-it-eugr | bench-agent-v2v2.0 | 22.2t/s | — | — | ~50,000 | fail |
| golden 256k/fp8 @ 22.2 tok/s — fill~50000 — bench-agent-v2 — tool_ok=False | |||||||
| 2026-06-21 | google-gemma-4-26b-a4b-it-eugr | bench-agent | 21.8t/s | 21.2 · 21.0 · 23.0 | 21.0–23.0 | @ 8k | — |
| agent bench avg 21.8 tok/s over 3 sessions × 3 turns (2023 tok in 93.1s) | |||||||