
Depth Pro 2 Shipped Without a Paper: What the Weights Show
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 358 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 139 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 80 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 111 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Scroll to the bottom of Apple's Depth Pro 2 model card and read the citation. Eight authors, a 2026 date, and an arXiv identifier that reads 261X.XXXXX. The Xs are not a redaction. They are a placeholder that nobody has come back to fill in, and they are the most honest summary of the release: Depth Pro 2, Apple's metric monocular depth model, exists as weights you can download today, and it does not yet exist as a paper you can read.
That gap is the story. Not a launch event, not a benchmark table, not a demo Space — the repository itself flags the Space as hidden until it goes public, and as of this writing it has not. What follows is everything that is knowable from the repositories and the documentation, kept separate from what is not yet confirmed.
What shipped, and on which dates
The dates are the part that is documented, because Hugging Face timestamps every repository it hosts. The model card repository, apple/DepthPro2, was created on 3 November 2026. The weights repositories followed two days later, on 5 November, with apple/DepthPro2-H among them. The Core ML export repository came on 9 November. The standalone focal-length estimator, apple/DepthPro2-focal, appeared on 19 November — the same day the code repository, apple-aiml-research/ml-depth-pro-2, went public on GitHub.
Read that sequence as an engineering timeline rather than a marketing one. The card, the weights, the Apple silicon export, the camera-calibration model and the code were staged over sixteen days, and the last piece to land was the one that makes the whole thing runnable by someone who is not inside Apple. Nothing about the order suggests a coordinated announcement; it suggests a team pushing artifacts as they cleared review.
The reception confirms it. The H+ weights repository has 40 downloads and the card repository 10 likes. The GitHub repository has a single commit behind it. For a model from the lab that produced Depth Pro, that is not a release curve — it is a release that has not been noticed yet.
The paper is a placeholder, not a preprint
Three separate places point at a technical report that does not exist. The model card's badge links to arxiv.org/abs/TODO. The GitHub README's badge links to arxiv.org/abs/261X.XXXXX. The BibTeX block in both says journal = {arXiv preprint arXiv:261X.XXXXX} and url = {https://arxiv.org/abs/261X.XXXXX}.
There is no Apple Machine Learning research page for it either — machinelearning.apple.com/research/depth-pro-2 returns a 404, and Apple's research sitemap lists depth-pro but not depth-pro-2. The project page the badges point at, apple-aiml-research.github.io/ml-depth-pro-2, is the one Apple-controlled surface that does resolve.
This matters more than it sounds like it should. A technical report is where the evaluation protocol lives, and without it the numbers the card reports cannot be checked against the numbers the baselines report. Everything quantitative below is therefore the vendor's own claim, reproduced without independent verification, and labelled as such.
Four sizes, plus a focal-length estimator
The family is unusually explicit about its own scaling. Apple publishes parameter counts and download sizes for all four members, which is more than most depth releases bother with:
• Small (S) — 24M parameters, 97 MB download, the one to start with on a laptop.
• Base (B) — 97M parameters, 386 MB.
• Large (L) — 334M parameters, 1.3 GB.
• H+ (H) — 872M parameters, 3.5 GB, and the default the command line downloads if you do not pass -m.
• Focal estimator — 23.5M parameters, 94 MB, a separate repository that loads alongside any size and predicts the camera's focal length when the photo or video does not carry one.
The naming is worth a note, because it will trip up anyone writing a config file: the repository is apple/DepthPro2-H but the command-line flag is -m depthpro2-h, while the card calls that size H+. The documentation uses both spellings within a few lines of each other.
The architecture underneath is a multi-view transformer followed by a per-frame DPT decoder. After the first third of the transformer's blocks, attention alternates between within-frame and across-frame, and the decoder assembles tokens from several blocks into feature maps at one thirty-second to one quarter of the input resolution. The README credits TIMM and DINOv3 as the code and model foundations.
The tradeoff chart is the only result
The card's entire results section is one figure and two sentences. The figure plots Boundary F1 against Jitter, the two metrics Apple chose, as a scatter of roughly two dozen single-frame and multi-frame depth models, with the four Depth Pro 2 sizes in orange and everything else in blue or green, averaged over the Sintel and Spring datasets. The two sentences say Depth Pro 2 outperforms prior work operating on a single frame or multiple frames on both, and that the top right is best.

The axes are labelled: Boundary F1 runs from 0 to 0.3, jitter from 0.1 down to 0. Reading positions off a scatter, the H+ marker sits near 0.29 Boundary F1 with jitter around 0.017, against roughly 0.235 and 0.095 for the original Depth Pro — a large gap on both axes. What the figure does not carry is a table, a per-dataset breakdown, any speed or memory measurement, or a comparison outside the two dozen models plotted. The card says the paper compares the sizes, including their speed and memory — and the paper does not exist yet.
That is not a reason to distrust the model. It is a reason to distrust any article that hands you a Boundary F1 for Depth Pro 2 quoted to three decimals, because the only place such a number exists is a scatter you can read to about a hundredth, with the evaluation protocol deferred to a report that has not been published. If you need the figure for a procurement document, you are waiting for the report.
Video is what makes it a sequel
The original Depth Pro was, in its own title, about sharpness and speed: Sharp Monocular Metric Depth in Less Than a Second, 2.25 megapixels in 0.3 seconds on a standard GPU. Depth Pro 2 keeps the sharpness and adds a second axis — temporal consistency — and the whole architecture change exists to serve it. Alternating attention within and across frames is how a transformer sees several frames as one scene instead of several images.

Two operating modes come out of that. Buffered mode, the default for video files, processes clips of 16 frames that overlap by four, which keeps depth steady where consecutive clips meet; --window and --overlap expose both numbers. Unbuffered mode predicts each frame the moment it arrives, from that frame and the 15 before it, and is the mode for a live camera. Apple is candid that the first frames of a stream have fewer predecessors to attend to, so their depth is less steady — a warm-up penalty stated in the documentation rather than discovered in production.
The rest of the video behaviour is a set of consequences worth knowing before you integrate. A video gets a single focal length for all its frames, taken as the median of the estimates over its first window. Memory grows with frame size and window length, not with video duration, so the documented remedy for a clip that will not fit is to shorten the window first and drop to a smaller model second, resizing the input only as a last resort. And sky pixels come back as infinite depth, or, especially on the smaller models, as a finite depth of several kilometres — the README tells you to clip the range before building a point cloud, which is the sort of advice that only exists because someone did not.
What is still unconfirmed
The honest list of open questions is short and specific.
• Independent evaluation — none. No third party has published a Boundary F1 or Jitter number for Depth Pro 2, and the card's own baselines are unnamed beyond "prior work operating on a single frame or multiple frames".
• Speed and memory — claimed to be compared in the paper, which is a placeholder. The only concrete runtime guidance is that H+ is 3.5 GB of weights and that S is the laptop option.
• The demo — the Space is linked inside an HTML comment, marked hidden until public. A demo existing and a demo being reachable are different claims.
• The name — "Depth Pro 2" appears as DepthPro2 in repository names, depth-pro-2 in the code package and the Core ML repository, and depth_pro_2 in the Python import. Search for all four.

Running it before the paper lands
The install path is deliberately boring, which is a point in its favour. You install uv, clone the repository, and run uv run depth-pro-2 predict -i assets/example.jpg -o out/ for a photo or the same command with an .mp4 for a video. The first run pulls H+ at 3.5 GB unless you pass -m depthpro2-s. The command picks the first available device — CUDA, then Apple silicon MPS, then CPU — and runs under bfloat16 autocast on GPUs, float32 on CPU. There is a Core ML export for Apple hardware and a minimal Mac demo app that needs macOS 15 and Xcode 16.
The output formats are where the practical detail lives. The command always writes a colour-coded depth image, red near and blue far. --save-png16 adds 16-bit greyscale PNGs for compositing tools, white near and black far, scaled to the depth range of the photo or of a video's first frame — so points that come nearer later in a video saturate. --save-npz adds the actual depth array in metres plus focallength_px, which is the one you want if you are doing anything numeric.
Two conventions will bite you if you skip the documentation. Depth is the distance along the camera's viewing axis, the z coordinate, not the straight-line distance from the lens. And inverse depth in Python is normalized by focal length: it equals focallength_px / (width * depth). Get the focal length 10 percent wrong and everything lands 10 percent too far away.
None of this is routed through us. Depth Pro 2 is research-licensed weights from Apple's own repositories, and if you want to run it you download it from Apple. What OrcaRouter covers is the rest of the pipeline around it — one API and one key for 200-plus hosted models at 0% markup over provider list price, with automatic failover across providers. When the depth stage is a research checkpoint with a placeholder citation and no release cadence, the reasoning and writing stages around it are the ones you want to be the boring, redundant part.
What to watch, if you are waiting rather than building: an arXiv identifier replacing the Xs, a research page appearing under machinelearning.apple.com, and the demo Space going from commented-out to linked. Any one of those would turn this from a weights drop into a release, and would give the chart above the per-model table it is missing.
