Skip to content
The Model Company

NewAnnouncing our $14 million Series A from DCM Ventures

Create smaller,
faster, better
models made
for your work.

Trained on your examples to do one job better, at a fraction of the cost of general-purpose AI. No machine-learning expertise required.

Describe the task in plain language, for example: route customer support tickets to the right team.

No sign-up to try it. Demo jobs are simulated and free — you'll see exactly what real training would do.

01Why smaller

A general model knows a little about everything. Yours knows your work.

The sample support router vs. a large general-purpose LLM

Illustrative

  • 470×

    smaller

    149M vs ~70B parameters

  • 188×

    faster

    13 ms vs 2.4 s at p95

  • 90×

    cheaper per call

    $0.036 vs $3.24 per 1,000 calls

And more accurate, not less: 92.1% vs 85.9% on the same 240 held-out tickets.

Illustrative: the LLM was prompted with the same 6 labels. Latency is model time on one vCPU; a network hop adds a few tens of milliseconds. See the comparison

Why it can be this small

It only has to recognize the topics in your work and give one of your answers — not know everything a general model knows.

General-purpose model: 96 sample topics, from sonnets to tax law, all weighted equally.

96 topics, equal weightunbounded outputs

What it needs to say

  • BillingFinance operations
  • Technical issueTier 2 support
  • Account accessIdentity & security
  • Shipping & deliveryLogistics
  • CancellationRetention
  • Feature requestProduct

02How it works

From a sentence to a model you own.

The platform handles the machine-learning decisions — data preparation, model selection, training, evaluation and packaging. You make the product decisions.

  1. Step 1: Describe it

    Say what the model should do and what it should hand back — one category, a set of tags, or named fields.

  2. Step 2: Show it

    Upload past examples as a CSV or JSONL file — or start with ours. We check your data before anything runs.

  3. Step 3: Own it

    We train small candidates against your bar for quality, speed and cost, then hand you the model to run anywhere.

03Evaluation

Know how good it is before you ship it.

Every model is tested on examples it never saw, checked against the bar you set, and compared with what you would otherwise use.

support-ticket-router · v3 · evaluation report
Illustrative

Against your requirements

92.1%

accuracy on 240 held-out tickets

Clears your 90% bar

  • Latency (p95)13 ms≤ 50 ms
  • Training spend$0.00≤ $25.00

Weakest label: Feature request at 78.0% F1 with 18 test examples. More examples of it would help most.

Compared on 240 held-out tickets

  • Your modelyours

    Long-context encoder · 149M, fine-tuned on your examples

    Accuracy
    92.1%
    Latency p95
    13 ms
    Cost / 1k calls
    $0.036

  • Large general-purpose LLM

    Prompted with your label definitions (zero-shot)

    Accuracy
    85.9%
    Latency p95
    2.4 s
    Cost / 1k calls
    $3.24

  • Small general-purpose LLM

    Few-shot prompt with 20 of your examples

    Accuracy
    79.5%
    Latency p95
    731 ms
    Cost / 1k calls
    $0.44

  • Keyword rules

    Hand-written keyword routing — a typical starting point

    Accuracy
    65.6%
    Latency p95
    1 ms
    Cost / 1k calls
    $0

From the sample support router. Your numbers depend on your data.

Open the full evaluation

04For AI agents

Agents are customers, too.

Anything a person can do here, an agent will be able to do through MCP, a REST API or the CLI — with explicit permissions and a budget it can't exceed.

  • Capability discovery

    A manifest and llms.txt describe what's possible and what it costs.

  • Structured schemas

    Every tool and endpoint is typed with JSON Schema.

  • Long-running jobs

    Return immediately, then poll or subscribe to webhooks.

  • Idempotency

    Retries with the same key never create a second job.

  • Spending limits

    Every job needs a hard cap; larger ones need human approval.

  • Artifact export

    Signed downloads with checksums and a model card.

Today, your agent can already help you build one. Paste this into it:

Read the agent reference
Planned · not live
tools/call · create_model_job
{  "tool": "create_model_job",  "arguments": {    "name": "Support ticket router",    "task": {      "description": "Route support tickets to the right team",      "template_id": "support_routing",      "output_kind": "single_label",      "labels": [{ "name": "Billing", "metadata": { "destination": "Finance operations" } }, "…"]    },    "dataset_id": "ds_8f2k1q",    "requirements": {      "min_accuracy": 0.9,      "max_latency_ms": 50,      "deployment": "hosted_api",      "max_spend_usd": 25    },    "mode": "training",    "idempotency_key": "router-2026-10-17-a"  }}

05Ownership

Yours to keep.

In this preview your model runs on our API. Downloading the weights to run it anywhere (your cloud, your laptop, a device in the field) is planned, not live yet. Sizes shown are for the sample support router.

  • 01ONNXPortable graph for CPU and GPU runtimesPlanned · not live
  • 02safetensorsRaw weights for Hugging Face toolingPlanned · not live
  • 03GGUFFor llama.cpp-style runtimes, when the base is a decoderPlanned · not live
  • 04DockerA container with an HTTP endpoint, ready to deployPlanned · not live
  • 05Hosted APIOr let us run it, priced per prediction$0.036 / 1k

06Questions

Good questions.

Start with one sentence.

Your first demo job is about a minute away. No account needed.

Describe the task in plain language, for example: route customer support tickets to the right team.