Sold by: Bioptimus
Deployed on AWS
The leading foundation model for computational pathology. H-optimus-1 turns H&E whole-slide images into task-agnostic embeddings for biomarker discovery, mutation prediction, and survival modeling, with state-of-the-art performance on 13 downstream tasks. 1.1B-parameter Vision Transformer pre-trained on 1M+ slides from 800,000+ patients across 50+ organ systems.
Overview
Use H-optimus-1 to build pathology models from H&E whole-slide images. Generate task-agnostic embeddings that preserve cellular and tissue-level context, then plug them into classifiers, survival models, or downstream heads for biomarker discovery, mutation prediction, patient stratification, tissue classification, cell counting, typing, and segmentation. Zero-shot and few-shot adaptation lets you fit a new task with linear probing or light fine-tuning instead of a full training run.
H-optimus-1 is a 1.1B-parameter Vision Transformer pre-trained on over 1 million H&E slides from more than 800,000 patients across 4,000+ clinical centers and 50+ organ systems (healthy and diseased tissues). On internal and public benchmarks, it delivers state-of-the-art performance across 13 downstream tasks on 15 datasets, including the public HEST benchmark.
Deploy H-optimus-1 as an Amazon SageMaker model package inside your AWS account. Run real-time inference via SageMaker endpoints or batch inference on S3. Whole-slide images stay in your VPC.
Highlights
- H-optimus-1 delivers state-of-the-art performance across 13 downstream tasks on 15 public and private datasets, including the public HEST benchmark; supports zero-shot and few-shot adaptation via linear probing or light fine-tuning with minimal task-specific training.
- H-optimus-1 is pre-trained on a proprietary cohort of billions of image patches sampled from over 1 million H&E slides representing more than 800,000 patients across 4,000+ clinical centers and 50+ organ systems, spanning healthy and diseased tissues.
- H-optimus-1 is a 1.1B-parameter Vision Transformer for H&E whole-slide images pre-trained with self-supervised learning at 0.5 MPP and 224x224 px tile resolution, producing 1,536-d task-agnostic embeddings that preserve cellular and tissue-level context.
Details
Sold by
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
Pricing is based on actual usage, with charges varying according to how much you consume. Subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimension | Description | Cost |
|---|---|---|
ml.g5.xlarge Inference (Batch) Recommended | Model inference on the ml.g5.xlarge instance type, batch mode | $540.00/host/hour |
ml.p3.2xlarge Inference (Batch) | Model inference on the ml.p3.2xlarge instance type, batch mode | $540.00/host/hour |
inference.count.m.i.c Inference Pricing | inference.count.m.i.c Inference Pricing | $0.001/request |
ml.g5.2xlarge Inference (Batch) | Model inference on the ml.g5.2xlarge instance type, batch mode | $540.00/host/hour |
Dimensions summary
H-Optimus-1 uses pay-as-you-go pricing with no upfront commitment. You pay by the hour for batch inference on three instance types: ml.g5.xlarge, ml.g5.2xlarge, and ml.p3.2xlarge. These options let you match compute power to your workload, so your cost depends on which instance you pick and how long it runs. A separate dimension bills by request count rather than host hours. Across all options, you deploy the model as a managed endpoint and generate feature embeddings from H&E whole-slide images. Your total cost scales with instance choice, runtime, or request volume.
Top-of-mind questions for buyers
Each host-hour dimension runs the model on a GPU instance in batch mode. The ml.g5.xlarge and ml.g5.2xlarge options differ in memory and compute size, while ml.p3.2xlarge uses an earlier GPU type. You pick the instance that matches your slide volume and speed needs.
The host-hour dimensions meter the time your instance runs, not the work it does. A deployed endpoint that stays running still accrues host-hour charges even between jobs. Shut the endpoint down after your batch finishes to stop software charges. Underlying AWS infrastructure fees may still apply separately.
The host-hour dimensions bill for instance running time, so cost grows with how long the endpoint stays up. The request-count dimension bills per inference request instead, so cost tracks the number of calls you make. Running time suits steady batch jobs; request billing suits intermittent, call-driven use.
docs.bioptimus.com
Helpful?
Vendor refund policy
Non-refundable.
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
v2.3.0 - Combined mode returns a tile embedding and its downstream output in a single forward pass. - Anonymized diagnostics endpoint (/diagnostics) and support bundles for reporting issues without sharing slide data. - Endpoint verification against the model's declared output contract, with typed error codes and remediation hints. - Coordinate-driven tile extraction (explicit tile list instead of a mask). Reliability - Slide-edge reads clamped to slide bounds; cuCIM and OpenSlide paths unified. - Server reports a terminal unhealthy state when all models fail to load. - Per-tile failures are isolated instead of aborting the whole slide. Compatibility - H-optimus-1 embeddings (1536-dim) and tissue segmentation unchanged. - Recommended instance: ml.g5.xlarge.
Additional details
Inputs
- Summary
H-optimus is a Foundation Model which accepts 224x224 pixel images at 0.5 MPP resolution and outputs an embedding feature vector of dimension 1536. The model accepts a single JSON object per request
- Limitations for input type
- Please note the input resolution should be 0.5 MPP (microns-per-pixel). Lower resolutions work less well
- Input MIME type
- application/json
https://github.com/bioptimus/h1-jumpstart/tree/main/data/input/real-time
https://github.com/bioptimus/h1-jumpstart/tree/main/data/input/batch
Input data descriptions
The following table describes supported input data fields for real-time inference and batch transform.
Field name | Description | Constraints | Required |
|---|---|---|---|
model_name | Selects the model to run. Must be "h1".
| String | Yes |
mode | Inference mode. Must be "embedding".
| String | Yes |
image_data | Base64-encoded PNG of a 224×224 RGB tissue tile extracted at 0.5 µm/px (MPP).
| String (base64)
| Yes |
slide_name | Identifier for the source whole-slide image.
| String | Yes |
x | Left edge of the tile within the source slide, in pixels.
| integer ≥ 0
| Yes |
y | Top edge of the tile within the source slide, in pixels.
| integer ≥ 0
| Yes |
width | Tile width in pixels. Must be 224.
| integer = 224
| Yes |
height | Tile height in pixels. Must be 224.
| integer = 224
| Yes |
patch_idx | Zero-based index of the tile within the slide.
| integer ≥ 0
| Yes |
tissue_ratio | Fraction of the tile area containing tissue (0.0–1.0)
| float [0.0, 1.0]
| No |
Resources
Vendor resources
Support
Vendor support
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
Multimodal foundation model for oncology. M-Optimus-1 turns a routine H&E slide into ~6,000-gene spatial expression readouts. Pre-trained on millions of proprietary H&E whole-slide images (50+ tissues) plus paired patient cohorts with aligned H&E, bulk RNA-seq, and spatial transcriptomics.
Organizations waste up to 40% of their AWS spend on underused EC2 and S3 resources, usually without realizing it. h‑Bar Solutions eliminates that waste with automated, data‑driven optimization that right-sizes compute, removes storage inefficiencies, and maintains performance. Customers choose us because we make cost reduction predictable, transparent, and painless without disrupting existing workloads or operations.
When teams shop for Managed Amazon SageMaker Operations, they worry about wasted spend, hidden complexity, and unreliable support. h‑Bar Solutions removes that fear by giving you transparent pricing, battle‑tested MLOps processes, and a dedicated expert team that keeps your models running flawlessly. You get predictable outcomes, faster deployment, and the confidence that your AI investment is actually delivering measurable value starting day one.
InvariMatch Nano by InvariVision detects pirated, duplicate, and modified video content with a false positive rate below 0.01% per image using proprietary frame-level image matching technology, optimized for cost-efficient 2-4 vCPU Amazon EC2 instances.
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.