← Compilation benchmarks

Gemma 4 26B A4B

@google/gemma-4-26b-a4b-it · NVIDIA B200 · May 2026 · vs SGLang

compiled model code size
96MB
compiled model code size
artifact pull, vs 25.8GB · 193,735 files
760MB · 5 files
artifact pull, vs 25.8GB · 193,735 files
CTTFT @p90 · Muna (bare metal)
6.3s
CTTFT @p90 · Muna (bare metal)
lower median CTTFT vs SGLang (Modal)
18×
lower median CTTFT vs SGLang (Modal)
TTFT @16k-token prompts
147ms
TTFT @16k-token prompts
tok/s across prompt categories
206–593
tok/s across prompt categories

All stats above are computed from the raw result files linked at the bottom of this page. See methodology for how each metric is defined and measured.

Artifact size

Everything that must be pulled onto a machine before it can serve a single token: the stock SGLang container image versus the entire Muna-compiled Gemma 4 artifact. Hover any slice to see what it is.

lmsysorg/sglang:latest

loading…

Hover a slice to inspect a file.

Muna-compiled Gemma 4

loading…

Hover a slice to inspect a file.

Cold-start Time to First Token

Cold-start Time to First Token: the time from a cold start to a user's first token, including container/artifact fetch, runtime init, model load, prefill, and first decode. Page cache was flushed before every bare-metal run.

loading…

Time to First Token

Warm time to first token across prompt lengths at batch size 1. Kernels are sourced from FlashInfer, TensorRT-LLM, and CUTLASS at compile time.

Time to First Token vs. Prompt Length

loading…

Decode throughput

Decode throughput across reasoning, math, code, knowledge, and summarization prompts at batch size 1, with speculative decoding via DFlash.

Decode Throughput by Prompt Category

loading…

Run configuration

Model
@google/gemma-4-26b-a4b-it
Weights
google/gemma-4-26B-A4B-it
Hardware
NVIDIA B200
Run date
May 2026
Baseline
lmsysorg/sglang:latest
Cold boots
5 per target
Host
RunPod B200 (kv85we196c0zw2)
Storage
local NVMe overlay (/), 8x SOLIDIGM SB5PH27X076T md0 RAID10
Cache eviction
posix_fadvise(POSIX_FADV_DONTNEED) per boot; /proc/sys/vm/drop_caches unavailable in container
TTFT rounds
50 per prompt length, 3 warmup, batch size 1
Throughput rounds
20 per category, 512 tokens, batch size 1

Raw data

Every chart on this page is rendered directly from these result files, including the raw per-run samples. Rerun the benchmark yourself with the reproduction script.

Run this compiled model.

Serve Gemma 4 26B A4B on Muna with an OpenAI-compatible endpoint, or deploy the same compiled model to your own compute.

$ muna deploy @google/gemma-4-26b-a4b-it --provider modal --gpu b200