Pricing

A flat annual licence. No per-token fees.

Modelsmith is priced as an annual software licence. The full platform at one price: the iterate loop, promotion state machine, MCP control-plane tools, CLI, and governance tooling. You own the hardware; we help you select the tier that fits your training cadence and model count.

No usage billing

Train, evaluate, and promote as often as your hardware allows. The licence does not meter inference calls or training steps.

No data egress

Your training data, your model weights, your evidence bundles. Nothing leaves the hardware you control.

No vendor lock-in

Promoted models are standard open-weights artefacts. Export them, serve them, or migrate away at any time.

Licence tiers

Three tiers. One pricing model.

All tiers include the full Modelsmith platform. The tier determines how many domain specialist lines you can run in parallel and the level of support cover.

Tier A

Starter

From £60,000/year

One to three specialist lines. The complete Modelsmith platform for a focused first deployment.

  • Up to three domain specialist model lines
  • Full iterate loop: eval, diagnose, post-train, promote
  • Six-state promotion state machine
  • CLI and MCP control-plane tools
  • Evidence bundle exports
  • Email support

Tier B

Popular

Growth

From £150,000/year

Four to eight specialist lines across adjacent workflows. Priority support and governance reporting.

  • Four to eight domain specialist model lines
  • Everything in Tier A
  • Priority support with faster response SLA
  • Governance reporting exports
  • Access to beta features

Tier C

Enterprise

Priced on scope

Unlimited specialist lines, dedicated success engineering, and procurement-compatible contracts.

  • Unlimited domain specialist model lines
  • Everything in Tier B
  • Dedicated success engineer
  • Custom SLA
  • Multi-year and volume pricing
  • MSA, DPA, and SOW templates

All prices are exclusive of VAT. Annual licences are invoiced on signature, net 30. A 10% prepayment discount applies when the Year 1 licence is paid in full on signature. Multi-year pricing is available at Tier B and above.

Design Partner programme

2026 cohort: 2 slots

Two partners. First cohort. Published terms.

A design partnership is a co-development relationship, not a discount scheme. We cap the 2026 cohort at two partners so each deployment gets the attention a first-cohort partner deserves: direct input into the roadmap, a named engineering counterpart, and a co-developed evidence bundle you can take to your governance team.

Pricing is published below. Payment is 50% on signature, 50% on Kickstart milestone acceptance within 30 days of delivery. Year 2 reverts to Tier B list.

Year 1 total for a design partner

Tier B licence, Year 1

Four to eight specialist lines. Reverts to Tier B list price from Year 2.

£125,000

Specialisation Kickstart

All five Professional Services phases bundled. Discounted from £155,000 à la carte.

£100,000

Year 1 total

Hardware is customer-procured, excluded. Excluding VAT.

£225,000

Professional Services

From first eval to production sign-off.

A Professional Services engagement covers the five deliverables required to take a domain specialist from definition to a production-validated model. Each phase is priced and scoped independently.

Engagements run 6 to 8 weeks of calendar time, delivered async-first with batched milestone reviews. Payment is 50% on signature and 50% on milestone acceptance within 30 days of delivery.

Specialisation Kickstart

£100,000 bundled

All five phases. Reserved for Design Partner engagements.

À la carte totals £155,000. The Kickstart bundle is the design-partner route and is included in the £225,000 Year 1 total described above.

  1. Eval design

    £30,000

    Define benchmark tasks, scoring rubrics, and acceptance thresholds for your domain. The eval suite becomes the standing gate for every subsequent promotion. We work with your domain experts to establish what good looks like before any training begins.

  2. Scenario generation

    £40,000

    Author synthetic training scenarios calibrated to your real workflow data distribution. Scenarios are structured as (input, target, rationale) triples that Modelsmith can ingest directly. No proprietary data leaves your organisation during scenario generation.

  3. Connector configuration

    £20,000

    Wire Modelsmith to your existing data stores, CI systems, and model registries. Includes configuration of the Modelsmith CLI against your hardware, validation of GPU access, and smoke-testing the full iterate loop end-to-end.

  4. First training campaign

    £50,000

    Run the first closed-loop specialisation cycle: baseline eval, diagnosis, post-training, and promotion review. We pair with your team throughout so they can operate subsequent campaigns independently.

  5. Production-validation handoff

    £15,000

    Sign off the specialist model against a representative live-traffic or held-out sample. Produces a complete evidence bundle: eval transcripts, promotion decision log, model card, and rollback procedure. The handoff document is the artefact your governance team approves.

À la carte pricing excludes VAT. Phases can be scoped independently or bundled as the Specialisation Kickstart for design partners.

Hardware guidance

More hardware means faster convergence, not more specialists.

Agentsia does not sell hardware. The configurations below are reference tiers to help you size your initial investment. A single specialist model line requires only an Entry configuration; additional hardware shortens the training cycle, it does not increase your model count.

Hardware is a one-off capital expense. All subsequent training campaigns run on the same node at no marginal cost.

Edge

£1,000 – 3,000

Models
Up to 4B parameters
Cadence
Weekly or ad-hoc

Single consumer-grade GPU or NPU. On-device inference for privacy-first workflows. Suitable for routing agents and lightweight classifiers.

Entry

£3,500 – 5,000

Models
Up to 14B parameters
Cadence
Overnight training runs

One Nvidia consumer-grade GPU. The reference configuration for a single-workflow specialist. Training completes overnight; morning promotion reviews are the norm.

Mid

£15,000 – 25,000

Models
Up to 32B parameters
Cadence
Several runs per week

Workstation-class Nvidia hardware. Supports multiple specialist lines and a weekly training cadence. The recommended starting point for Tier B licences.

Enterprise

£50,000 – 150,000+

Models
70B+ parameters
Cadence
Daily or continuous

Multi-GPU rack node. Daily training cadence across many specialist lines. Designed for Tier C deployments with high-volume eval requirements.

Hardware costs are indicative list prices for new consumer-grade Nvidia hardware as of 2026. Contact us for guidance on specific configurations before purchasing.

Synthetic POC

£25,000. Credited against your licence on conversion.

The Synthetic POC is a fixed-fee proof-of-concept engagement. We generate domain-representative training scenarios together, run a complete specialisation cycle on your hardware, and hand over a validated model with a full evidence bundle.

Your production data never leaves your environment. The POC uses only synthetic scenarios calibrated to your domain. The £25,000 fee is paid upfront and credited in full against your Year 1 licence if you convert within 60 days.

What you get

  • Trained and promoted specialist model
  • Full evidence bundle with eval transcripts
  • Promotion decision log
  • Model card and rollback procedure
  • Recommendation for licence tier and hardware

Frequently asked

Do we need to share our production data?
No. The synthetic POC runs entirely on scenarios we generate together from your domain specification. You describe the workflow; we author the training scenarios. No proprietary data leaves your organisation.
How long does it take?
Three to four weeks from kickoff to evidence bundle, assuming hardware is provisioned. The timeline depends on domain complexity and the availability of your subject-matter experts for scenario review.
What is the output?
A fully trained and promoted specialist model, plus a complete evidence bundle documenting every evaluation gate it passed. The bundle is the same format your governance team will review in production.
Can we run the POC on our own hardware?
Yes. We can deploy to your hardware from day one, or work on a temporary reference configuration if your hardware procurement is still in progress. The POC is hardware-agnostic.
How does the licence credit work?
The £25,000 POC fee is credited in full against your Year 1 licence if you sign a licence agreement within 60 days of POC completion. Convert to Tier A, Tier B, or the Design Partner programme and the POC is effectively free; walk away and the fee covers our work.
Is the POC required to enter the Design Partner programme?
No. Design partners can skip the POC and go straight into the Specialisation Kickstart. The POC is the lower-commitment route for teams that want to see a domain-representative result before committing to a Year 1 contract.

Year 1 examples

Two worked examples for Year 1.

The first example is a self-serve deployment on Tier A with a starter Professional Services scope. The second is the Design Partner programme Year 1, capped at two slots for the 2026 cohort. Both assume an Entry hardware configuration purchased directly from Nvidia.

Self-serve route

Tier A + starter PS + Entry hardware

Modelsmith Tier A licence

12 months

£60,000

Annual

Starter PS: eval design + first campaign

Two of the five Professional Services phases

£80,000

One-off

Entry hardware

One Nvidia consumer-grade GPU, one-off capital

~£4,250

One-off

Year 1 total

Excluding VAT

~£144,250

Illustrative. Starter PS combines eval design and the first training campaign; connector configuration, scenario generation, and handoff are priced separately.

Design Partner route

2026 cohort, 2 slots available

Tier B licence, Year 1

Four to eight specialist lines. Reverts to Tier B list Year 2.

£125,000

Annual

Specialisation Kickstart

All five PS phases bundled, saving £55,000 vs à la carte

£100,000

One-off

Entry hardware

One Nvidia consumer-grade GPU, one-off capital

~£4,250

One-off

Year 1 total

Excluding VAT

~£229,250

Payment is 50% on signature and 50% on Kickstart milestone acceptance within 30 days. Hardware is customer-procured and excluded from the 50/50 structure.

From Year 2

Hardware is a one-off capital expense. Professional Services are optional after the first engagement. From Year 2, the annual cost is the licence renewal at list.

Further specialists

Additional domain specialist lines run on the same hardware. There is no per-specialist hardware cost; your existing node handles multiple lines in sequence.

Get started

The right place to start is a 30-minute scoping call.

We will talk through your domain, your hardware situation, and your governance requirements. You will leave knowing which licence tier fits, whether Professional Services make sense for your first engagement, and what a realistic Year 1 timeline looks like.