PLAN YOUR DEPLOYMENT

Inference requirements
questionnaire

Tell us what you need to run. We'll use your answers to scope compute, pricing and service targets around your workload.

No account neededEstimates welcomeTechnical sections optional
01

Your team & workload

Start here. The remaining sections are optional; estimates and "not sure yet" are welcome.

A date, quarter or rollout window.

Describe the application, who uses it and what you want to improve.

Measured or estimated? Include the observation period, time zone and expected growth.

02

Models & quality

Tell us what needs to run, and how you judge a useful answer.

Include public model links, private fine-tunes, parameter size and any licence restrictions.

For each model; distinguish supported context from the context you actually need.

Reference dataset, metric and target score; tolerance for quantisation or model changes. Do not upload datasets.

03

Input & output lengths

Input includes the system prompt, conversation, retrieved context and tool results. P99 is the length at or below which 99% of requests fall.

Input tokens per request

Use token counts, not characters.

Output tokens per request

Include reasoning tokens where your model exposes them; explain counting differences below.

List separate profiles for different models or workloads, and note whether input and output peaks coincide.

04

Traffic & capacity

Tell us about both ordinary traffic and bursts. For rate metrics, specify the sampling window in the notes.

Simultaneous in-flight requests

Requests running at the same time.

Requests per minute

Measured or estimated over one-minute windows.

Input tokens per minute

Keep input and output token rates separate.

Output tokens per minute

Total output across all concurrent requests.

0 = all real-time; 100 = all batch.

Expected volume in 3–12 months.

Steady, bursty or scheduled peaks? Include peak hours/time zone, burst duration, batch deadlines and measurement windows.

05

Latency & service targets

These are requirements for evaluation, not an agreed SLA. State the client location, load and sequence lengths under which each target should hold.

Time to first token (seconds)

Elapsed time from sending a request to receiving its first output token, including queueing and network time if measured by the client.

Output speed per active request (tokens/second)

Decode speed after the first token. For a tail target, specify a lower percentile (such as P1), rather than P99.

End-to-end request latency (seconds)

From request submission to complete output.

Client-side or server-side? Streaming? Required lower percentile for speed, inter-token latency, concurrency and input/output profile. Describe an acceptance test.

Desired monthly availability, acceptable error/timeout rate, maintenance windows, support hours, incident response and recovery objectives.

06

Prompt caching

Complete this if repeated prefixes or cache behaviour matter to your workload. Leave blank if unknown.

Percentage of requests that reuse a cached prefix.

For example, 5 minutes or session duration.

Size limits, expiry and isolation between customers.

Empty cache: upper limit in seconds, percentile, sequence lengths and concurrency.

Steady-state cache: upper limit in seconds, percentile, sequence lengths and concurrency.

07

Deployment & integration

Describe your preferred setup. Deployment options and locations are subject to qualification and a separate quote.

Managed API, dedicated compute, root/bare-metal access, existing infrastructure, or still exploring.

Required countries/regions, tenant isolation, retention, private networking and compliance needs. Include NDA requirements before sharing further details.

Compatibility, streaming, tools/function calling, authentication and network/storage connections.

Advanced infrastructure details Optional

Current or intended phase separation, runtimes and compatibility requirements.

Formats, transfer mechanisms and accelerator/host/storage placement.

Bandwidth, latency, Ethernet/fabric, storage and connectivity constraints.

Rack space, power, cooling and readiness dates. This is scoping information, not a confirmed customer-site installation offer.

08

Budget & commercial expectations

Share the constraints that will help us scope a useful proposal. Include currency; leave commercial details blank if you prefer to discuss them privately.

For example, USD, SGD or EUR.

Provider/service, hardware if relevant, pricing tier, discounts and what the comparison includes.

09

Multimodal & structured output

Skip this section for text-only workloads without structured output requirements.

Image dimensions and formats, images per request, request share and other input modalities.

Resolution, frame rate, clip or audio duration, volume and latency requirements.

Share of requests, JSON schema/grammar constraints, tool use and acceptable validation failure rate.

10

Anything else?

Tell us about dependencies or requirements the other sections do not cover.

Custom service terms, procurement deadlines, evaluation process or next steps.

Submit your requirements

Send what you know. We'll review your requirements and follow up at the email address above.

Please don't include prompts, credentials, customer data or confidential model weights. Your contact details and answers are sent to Formspree to deliver your enquiry to Actaserve. Privacy notice.

Answers are not automatically saved. You can download a copy before leaving this page. Enquiries are retained until an authorised operator deletes them; request deletion at hello@actaserve.com.