`google/gemma-4-26B-A4B-it`

apache-2.0 1 launch

image-text-to-text · 25.81B parameters · 48.1 GiB of weights at revision 4d7ae498. Runs on a single GPU when the source verified it.

download the card performance the weights on HF

Did this card launch? The count is real — it comes from launches the kit recorded against this path, not from a widget anywhere.

1checkpoints

1published launches

1GPU minimum

2named gaps

The identity, read from the HF API at the revision the card pins.
weight path	`google/gemma-4-26B-A4B-it`
revision	`4d7ae4984b7db7de8f8457170b3f1a419ee76d52`
license	apache-2.0
gated	no
parameters	25.81B
weight bytes	48.1 GiB
modality	image-text-to-text
HF downloads	6.1M at capture
last modified	2026-07-20T16:42:12.000Z

The checkpoints.

Every servable variant in or beside the repo. resident_gib is what lands in VRAM after load; host_gib is what stays on the host and never touches the card.

name	format	location	resident	host tier	gated
`root`	bf16	repo root	48.1 GiB	—	no

root — 2 weight file(s)

The published launches.

One entry per engine × checkpoint with a command somebody published. Each cites the page it was read from — the card copies it, it does not invent it.

vllm `root`

vllm/vllm-openai:latest

argv — what a wrapper execs

vllm serve google/gemma-4-26B-A4B-it --tensor-parallel-size 1 --enable-auto-tool-choice --tool-call-parser gemma4 --chat-template examples/tool_chat_template_gemma4.jinja --reasoning-parser gemma4

as the source printed it

vllm serve google/gemma-4-26B-A4B-it \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice \
  --tool-call-parser gemma4 \
  --chat-template examples/tool_chat_template_gemma4.jinja \
  --reasoning-parser gemma4

validated on: h100, h200, mi300x, mi325x, mi355x, trillium, ironwood, xeon6, dgx_spark_gb10, dgx_station_gb300, rtx_pro_6000, rtx_5090, arc_pro_b60, arc_pro_b70
source: vllm-recipes

Every floor a box must clear.

A launch that fails on the third requirement wasted the first two. These are the requirements, as data — the fit gate reads them instead of a paragraph.

The hardware block, verbatim from the card.
GPU models	h100, h200, mi300x, mi325x, mi355x, rtx_5090, rtx_pro_6000
architectures	blackwell, cdna3, cdna4, hopper
min GPUs	1
TP rule	widest published TP is 1
min VRAM / GPU	—
min total VRAM	—
min disk	108 GiB
min host RAM	—
interconnect	not required
CUDA min	—
driver min	—
context tokens	—

The 2 things this card could not answer.

First-class, not swallowed. Every entry is a field a source did not answer — which makes it a launch nobody has done, and therefore a unit of work rather than a defect.

envs.sglang no SGLang cookbook page matched this repo

hardware.npu recipes name non-GPU accelerators too: arc_pro_b60, arc_pro_b70, dgx_spark_gb10, dgx_station_gb300, ironwood, trillium, xeon6 (not carded)

The file.

The card is a YAML document with a published JSON Schema. View it, download it, diff it against the next capture — and point one at it.

download

launch-card.yaml card.json the schema

Same bytes the collector wrote. The JSON is the same card, for a machine that would rather not parse YAML.

with one

one launch --card google/gemma-4-26B-A4B-it/launch-card.yaml

one follows the argv above, checks the floors before the paid clock, and on a failure changes exactly one field and retries — recording every attempt against this path.

Where every field came from.

The card is assembled, not authored. These are the sources, so a reader can check the claim rather than trust it.

source	url
hf-api	`https://huggingface.co/api/models/google/gemma-4-26B-A4B-it`
vllm-recipes	`https://recipes.vllm.ai/google/gemma-4-26B-A4B-it.json`

Captured 2026-10-31T17:28:30Z by hf-api + vllm-recipes + sglang-cookbook. Schema onefrontier.launch-card/v1.