Depth Anything 3 provider check: who serves DA3 today

As of 2026-10-26, one host sells Depth Anything 3 calls at a fixed price per image: WaveSpeed AI, at 0.005 an image, read on that date. Replicate carries community lanes that bill per GPU second. Hugging Face serves none of the ten DA3 tiers through its providers, in the Hub's own records. vast.ai has no Depth Anything 3 template, because it rents machines rather than serving models. On this platform the small tier is served under the id da3-small, on the public list since 2026-10-06. No other tier is on the public list. The publisher itself ships weights, code, and a free demo page. It ships no API of its own. All money figures below are in US dollars, published by the named seller on the date named.

Last verified: 2026-10-26

What it is

Depth Anything 3, or DA3, is the 2025 release of the Depth Anything family from a ByteDance Seed team. It reads any number of images, with or without known camera poses. It returns a depth map for each view, a confidence map, and the camera parameters. The paper calls it recovering the visual space from any views, submitted November 13, 2025. The repository and the Hub organization publish ten tier models. A sibling streaming variant handles ultra-long video. It needs under 12 GB of GPU memory, using a sliding window.

The tier table, from the repository's own model list, read 2026-10-26:

One row carries a real conflict. It sits inside the publisher's own two sources. The repository table marks DA3-LARGE-1.1 as CC BY-NC 4.0. That tier's own Hub card says Apache 2.0. Both were read on 2026-10-26. This family has changed license terms between releases before. The two-minute rule survives any conflict. Read the license block on the exact tier page you download from, on the day you download it. The tiers where both sources agree on Apache terms are small, base, metric, and mono. Those are the safe shipping set for a product today. The wider license question across robot model families has its own page, the GR00T license guide.

The name search mixes spellings and meanings, so the map earns its place. The publisher writes DA3. The public types V3. They are the same generation. The word replicate in a query is two questions at once. One is the host named Replicate. The other is the ordinary verb. This page answers the host in its census and the verb in its self-host paths. One more thing. The phrase depth anything no api found appears in Google's own autocomplete as a learned suggestion, read 2026-10-26. People typed it enough times for the engine to learn it. That is the demand signal this page exists to answer.

The posture on this platform, with its date: the small tier of this family is served here under the id da3-small, live since 2026-10-06, metered per image, with its own page. No other DA3 tier is on the public list as of 2026-10-26. This page promises nothing about future listings. The models index is the live catalog. The vision route guide documents the call contract the platform serves.

Use cases

Readers who typed depth anything v3 api want somebody else's GPU to turn images into depth. As of 2026-10-26 there are four real paths, and they have very different price units.

The first path is the priced per-image lane, and it is the newest fact in this census. WaveSpeed AI sells Depth Anything 3 image calls at a fixed 0.005 per image, read 2026-10-26. Colormaps and output formats add nothing to the bill. Its video lane prices at 0.005 per billed second. The billed length starts at three seconds and stops at six hundred. Depth is normalized across the whole clip, not frame by frame. For a video pipeline that needs temporal stability, that whole-clip normalization is the product. A fixed per-image price has one more virtue. You can multiply it by a counter before you spend anything. It was the only fixed per-unit DA3 price this research found on any host, as of the date above.

The second path is the community lanes on Replicate. One community publisher, vufinder, serves three DA3 variants there. They are the metric lane, a panorama variant of it, and the mono lane. All three bill 0.000975 per second on an L40S machine, read 2026-10-26. The median run prices, in the same order: 0.020 for the metric lane, nothing published yet for the panorama lane, and 0.00098 for the mono lane. A second publisher, david20321, serves the metric-large variant at 0.000225 per second on a T4 machine, median run 0.0024. The publisher of the model has no presence on Replicate. Two cautions come with these lanes. First, the unit. A per-second price only becomes a per-image price after you measure your own batch on your own images. Second, the license. Each lane names its tier. The metric and mono large tiers carry Apache terms. The giant tiers do not. Read the lane's own page before you ship its output. The older generations of the family ride the same marketplace at the same per-second shape. That is how the cheap-lane habit formed. It is also how the non-commercial traps of the older large tiers caught so many buyers.

The third path is the negative census, and it needs dates to be worth anything. On Hugging Face, every one of the ten DA3 tier records carries an empty provider list. That is the Hub's own model-record data, read 2026-10-26. The Hub distributes this family heavily: 211,716 downloads for the small tier alone. It serves none of it. Roboflow's hosted model docs list Depth Anything V2 and no V3 page. The V3 address returns not-found. fal's depth-anything-v3 address returns its own not-found page. DeepInfra's catalog carries no depth category at all. The multi-provider router listed 464 models that day. No depth model was among them. The two newest robot hosts, abstract.inc and roboticinference.com, were re-read the same day. The first keeps its model list behind sign-in and names no depth model in its public material. The second lists 24 robot action policies, with no depth model among them. Free of charge, the publisher's own demo Space, with 484 likes, takes image uploads in a browser. A demo is not a price. A download counter is not a service. The big-host survey across the whole robot model class lives in do the big hosts serve robot models.

The fourth path is rental plus self-host, and it is the working path for volume. vast.ai rents GPU machines with per-second billing. Its live table, read 2026-10-26 and stamped as updating hourly, showed 24 GB RTX 3090 boards from 0.12 an hour, median 0.21. The same table showed 24 GB RTX 4090 boards from 0.13 an hour, median 0.51. There is no Depth Anything 3 template in its library. The model is yours to install. The publisher's API documentation is the install path: one python class, one model id, one inference call. Self-host tooling grew fast in 2026. A ComfyUI node, ComfyUI-Depth-Anything-3, reached version 1.0.0 on 2026-10-02. It carries the small, base, mono large, and metric large tiers, with automatic downloads. An ONNX port exists for runtimes that want it. A hardware vendor wiki publishes a Jetson deploy guide with ROS2 and TensorRT for robot builds. The smallest tier needs no GPU at all. That is why ports of it keep appearing for edge runtimes. For depth on a moving machine, the pairing guide is robot vision segmentation and depth. The cost side of hosted versus onboard has its own page, the DA3 small pricing guide.

API usage

Three calls matter here. The first reads a live catalog. Listing state is a fact you read, never a fact you assume from a blog post. The second is the publisher's own pattern, the way to run any tier on a machine you control. The third is this platform's vision route, which serves the da3-small id today.

curl

Here is the catalog check, one request:

curl -sS "https://api.ironstratum.com/v1/models" \
  -H "Authorization: Bearer $KEY"

The answer lists every model id this platform serves, on the day you ask. As of 2026-10-26, da3-small is in it and no other DA3 tier is. Repeat the call later and you learn whether that changed. You will not need to trust any page's date, including this one's.

The vision call itself, against the served id:

curl -X POST "https://api.ironstratum.com/v1/vision" \
  -H "Authorization: Bearer $KEY" \
  -F "model=da3-small" \
  -F "file=@scene.jpg"

Metering is per image, from the count of images in the request. Rates for listed models sit on the pricing page, never in prose on pages like this one.

python

The working self-host pattern for any tier is the publisher's own, from the repository's API documentation:

from depth_anything_3.api import DepthAnything3

# On the GPU machine, after the repo is installed.
# Any tier id from the repository table works here.
model = DepthAnything3.from_pretrained(
    "depth-anything/DA3METRIC-LARGE"   # Apache 2.0, metric output
).to("cuda")

prediction = model.inference(
    ["frame_0001.jpg", "frame_0002.jpg"],  # one image or many
)

depth = prediction["depth"]        # per-view depth maps
conf = prediction["conf"]          # per-view confidence

This is the publisher's pattern, not a rewrite. The inference call takes one image or a list. It accepts known camera intrinsics and extrinsics when you have them, which improves consistency across views. For metric output from the metric tier, the repository gives the recipe in one line. Metric depth equals the focal length in pixels, times the raw output, divided by 300. The same class loads every tier in the table above. Your license choice and your hosting choice stay separate decisions.

http

The raw request to this platform's vision route, the contract the da3-small id serves under:

POST /v1/vision HTTP/1.1
Host: api.ironstratum.com
Authorization: Bearer $KEY
Content-Type: multipart/form-data; boundary=frame

--frame
Content-Disposition: form-data; name="model"

da3-small
--frame
Content-Disposition: form-data; name="file"; filename="scene.jpg"
Content-Type: image/jpeg

<the image bytes>
--frame--

Every field, error code, and limit lives in the run-vision reference, and the vision route guide walks the route end to end. A request naming an id that is not on the public list returns a named not-found error. That is the honest answer, and it is what the platform's own docs promise.

Benchmarks

Nothing in this section is this platform's measurement. A lane with no listing has no numbers of ours to publish, and none will be invented. Every figure carries its owner and its date.

The family claims come from the paper, read on its arXiv page 2026-10-26. DA3 beats VGGT, the prior multi-view state of the art, by an average of 44.3 percent in camera pose accuracy. It beats it by 25.1 percent in geometric accuracy. It matches Depth Anything V2 on detail and generalization, and beats V2 on monocular depth. Training used public academic datasets only.

The tier trade comes from the paper's speed table, on one A100, scenes of 32 images at 504 by 336. The small tier runs 160.5 images per second. It holds about 4,000 views in memory. It is the fastest and roomiest tier. The giant tier runs 37.6 images per second and holds about 900 to 1,000 views. On camera pose accuracy scored as AUC30, the small tier scores 71.9 on ScanNet++. The giant scores 98.1 on the same set. Read that as the honest trade. The small tier buys speed, memory headroom, and Apache terms. It gives up pose precision against its bigger siblings. For per-frame robot depth that trade is usually right. For survey-grade geometry it is not. The giant tiers that win those scores carry non-commercial terms.

The demand side, from the Hub's own records on 2026-10-26: 211,716 all-time downloads for DA3-SMALL, and 47,544 for DA3-LARGE-1.1. That is what an under-served family looks like from the outside. Heavy download traffic. One priced per-image lane. Community per-second lanes. No big host.

The community record comes from the repository's issue tracker, read 2026-10-26. It carries the rough edges worth knowing before you commit. Buyers asked for commercial-use permission. They asked for ONNX and conversion scripts. They asked for a mobile build. One complaint thread says DA3-SMALL trails the older V2 small tier on some scenes. Other reports mention artifacts and smoothness problems in hard scenes. None of that is a reason to skip the family. It is the difference between a paper table and a Monday morning.

Getting started

1. Read the primary sources. The paper, the repository with its model table and API documentation, and the Hub organization with the ten tier cards. Every identity fact on this page traces to one of them.
2. Pick the tier with the license table open. Small, base, metric, and mono carry Apache terms in both of the publisher's sources. The large-1.1 row is in conflict between them, dated above. Decide before you build, not after.
3. Match the path to your volume. A few hundred images a month: the priced per-image lane, or this platform's da3-small, metered per image. All day every day on a moving machine: a rented 24 GB board and the publisher's python API, or the ComfyUI node for video work.
4. Check the live list, not this page. Create an account on the signup page, make a key, and read the models index the way the curl tab shows. The docs pricing page holds the rates for listed models. The small tier's own page is da3-small, its cost walkthrough is the DA3 small pricing guide, and the vision category page holds the family this model belongs to.