On this page
DigitalOcean added NVIDIA Nemotron 3.5 Super VL to serverless inference on October 15, 2026. If your application turns screenshots or scanned pages into structured records, it is a new option worth evaluating. A low token price alone does not establish whether it can replace your current extractor.
This guide gives you a small, repeatable evaluation before you move customer documents. The request template is adapted from documentation; I have not run this model, measured its accuracy, or benchmarked its latency.
Disclosure: Tylor.nz has a DigitalOcean affiliate relationship. This article uses direct documentation links and contains no commission-bearing signup link.
What is actually available?
The October 15 release note announces serverless availability. It does not explicitly assign this model a GA or preview label.
DigitalOcean's model catalog lists the identifier nemotron-3.5-super-vl, Chat Completions and Responses support, and text, image, document and GUI-screenshot inputs. Audio is excluded. Dedicated inference is not marked supported. EU data residency is unavailable.
The catalog displays 1,048,576 in both the context-window and maximum-output columns. Do not budget for a million input tokens plus a million output tokens in one request. NVIDIA's model card explains that input and output share the configured context window and describes documents as page images with accompanying text.
Start with page images. The sources reviewed do not establish a DigitalOcean-specific direct PDF-upload contract, model-specific image byte limit, or page-count limit. NVIDIA capabilities are not automatically a guarantee that every input type works through DigitalOcean's deployment.
Decide whether the evaluation is worth doing
This is a plausible candidate when layouts vary and you need to extract both text and visual relationships: a total next to its label, an item aligned with the correct column, or an application's visible status.
Keep your existing approach in the comparison:
- For digitally generated PDFs with a reliable text layer, try a deterministic parser first.
- For fixed-layout scans, compare OCR plus rules. It may be easier to validate and operate.
- For variable screenshots, compare your current vision model using exactly the same inputs and scoring rules.
Do not begin a production migration if you require verified EU residency, a model-specific throughput commitment, or a retention policy you cannot reconcile with the caching behavior below. Resolve those requirements before sending real records.
Build a fixture with answers you already know
Create 30 synthetic page images, using invented names, amounts and identifiers. A useful first set is ten clean pages, ten with small text or skew, and ten with missing or ambiguous fields. Include look-alike characters such as 0 and O, multiple totals, and a page with no requested value.
Save each image alongside an expected record. For an invoice-like fixture, use four fields: invoice number, currency, total and whether human review is required. Define in advance what counts as a correct missing value. A model should not invent a plausible total when the image cannot support one.
This is a smoke test, not a representative accuracy study. Thirty examples can expose obvious failures; they cannot substantiate a production accuracy percentage for all your customers.
Freeze the fixture, prompt and output rules before comparing models. Changing all three after inspecting results makes a cheap-looking winner difficult to trust.
Use a documented image request shape
DigitalOcean's multimodal guide documents Chat Completions requests with text and image_url parts. It accepts HTTPS image URLs or base64 data URIs and lists PNG, JPG, JPEG and WEBP. Its worked example uses another Nemotron model; the template below adapts the model identifier and has not been executed.
Before running it, review current billing and terms yourself. Serverless inference requires credits or a prepaid balance. Use an existing model-access key with the appropriate permissions; do not paste it into source control, screenshots or a support ticket.
Prepare request.json, replacing the image placeholder with a base64 data URI for one synthetic PNG:
{
"model": "nemotron-3.5-super-vl",
"messages": [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Extract invoice_number, currency and total from this synthetic invoice. Return exactly one JSON object with invoice_number (string or null), currency (string or null), total (string or null), and needs_review (boolean). Preserve printed values. If any requested field is absent, unreadable or ambiguous, return null for that field and set needs_review to true; otherwise set needs_review to false. Do not guess."
},
{
"type": "image_url",
"image_url": {
"url": "data:image/png;base64,REPLACE_WITH_SYNTHETIC_IMAGE_BYTES"
}
}
]
}
],
"max_completion_tokens": 1024
}
Then send the file from a shell where your existing key is already available as MODEL_ACCESS_KEY:
curl --fail-with-body \
https://inference.do-ai.run/v1/chat/completions \
-H "Authorization: Bearer ${MODEL_ACCESS_KEY}" \
-H "Content-Type: application/json" \
--data-binary @request.json
The current Chat Completions guide recommends max_completion_tokens; older multimodal examples use deprecated max_tokens. Treat 1,024 as a starting output budget. Inspect finish reasons and truncation before adjusting it.
Asking for JSON in a prompt is not schema enforcement. Parse the response, validate required keys and types, reject extra prose, and check extracted values against the image. A successful HTTP response does not mean the record is correct.
Score useful records, not attractive answers
For every fixture, record the model, observation time, response status, finish reason, reported usage, elapsed time, parsed fields and manual verdict. Score each original page once, including pages that end in network failure or malformed JSON. Track individual retry attempts separately for spending and end-to-end latency.
Separate these outcomes:
- Correct and usable without editing.
- Correctly sent for human review because evidence was missing.
- Incorrect but presented as usable.
- Operational failure or invalid output.
The third category is especially important. A polished, incorrect amount can be more expensive than a request that clearly fails. Decide which fields must be exact before any output enters your application, and retain human review for consequential records.
Run the same fixture through your current approach. If you retry a failed candidate request, apply the same retry policy to the baseline and include all attempts in cost and latency. Keep a separate, untouched fixture for the final check after prompt changes.
Budget from measured tokens
DigitalOcean's pricing page lists these rates on October 16, 2026, in USD per million tokens:
- Uncached input: $0.30
- Cached input: $0.06
- Output: $0.65
For illustration, 10,000 uncached input tokens and 1,000 output tokens cost $0.00365. One thousand requests with exactly that usage cost $3.65. This is arithmetic, not an estimate for a thousand pages: image processing, prompt length, output length and retries change usage.
Use observed token categories without double-counting cached tokens. Then calculate total inference spend divided by accepted, usable records. Track human-review time separately. A cheaper request can still produce a more expensive workflow if it needs more corrections.
Do not assume a cache discount. Measure cache hits for your actual request pattern and report cold and warm runs separately.
Review caching before customer data
DigitalOcean's privacy documentation says DO-hosted requests stay on its infrastructure, are not sent back to the model creator, and are not used by DigitalOcean for model training. That is not a blanket claim that every part of your workflow retains nothing.
Signals changes the storage picture: when enabled, raw trace content is processed and stored in a US region; redaction is for Console display and is not a compliance guarantee. Your application, logs and browser can also retain content. Keep the evaluation synthetic and leave optional trace collection out until you have reviewed its implications.
DO-hosted prompt caching is automatic and cannot be opted out. It retains internal key/value state rather than prompt text, has no configured expiry, and uses least-recently-used eviction. Configurable prompt_cache_retention is unsupported. Caches are isolated by customer account; prompt_cache_key supports additional separation between tenants within an account.
If your policy requires an explicit maximum retention period or forbids retained inference state, do not infer compliance from a general no-training statement. Resolve that mismatch before transmitting customer material.
Check the operational fit before switching
DigitalOcean's limits page says serverless tier maximums do not guarantee throughput, latency or availability. Requests may be queued, throttled or rejected. DO-hosted models are excluded from batch inference.
The availability matrix does not provide a serverless row identifying this model's serving regions. Do not borrow a dedicated-inference region list as evidence for this endpoint.
Test your real concurrency requirement with a bounded, explicitly budgeted run after the small fixture passes. Use timeouts, limited retries, a review queue and a fallback path. Avoid an unlimited retry loop when capacity or validation fails.
My decision rule: keep the current extractor unless this candidate meets your predeclared correctness and abstention requirements, has acceptable measured cost per usable record, and fits your data-handling requirements. If it passes, trial a narrow, reversible workflow first. The launch gives you another option; your own fixture determines whether it earns a place in the application.
Sources checked October 16, 2026. No paid inference requests or hands-on model benchmarks were performed for this guide.