NEW

Language, engineered into poetry.

Language, engineered into poetry.

Language, engineered into poetry.

Machine Poetry builds expressive language models that reason like engineers and write like poets — deployed in minutes, trusted in production.

lyra-3 · session 0x7f2a

● Streaming

› Write a release note for our payments API in the voice of Rilke.

tone: lyrical

facts: verified

38 ms

Token stream

Coherence

0.991

Tokens / sec

1,240

Hallucination

0.02%

Trusted by teams who ship language at scale

[01] — Manifesto

We believe machines should not merely answer. They should compose. Every model we ship is tuned for clarity, cadence and consequence — so your product speaks with intent, not autocomplete.

We believe machines should not merely answer. They should compose. Every model we ship is tuned for clarity, cadence and consequence — so your product speaks with intent, not autocomplete.

We believe machines should not merely answer. They should compose. Every model we ship is tuned for clarity, cadence and consequence — so your product speaks with intent, not autocomplete.

Ines Varga

Co-founder & Chief Scientist

[02] — Models

Three models. One unmistakable voice.

Pick the mind that fits the moment. Every Machine Poetry model shares one alignment stack, one API and one obsession with craft.

01 / 03

Flagship

Lyra-3

Lyra-3

Lyra-3

Our most expressive model. Long-form reasoning, brand-perfect tone and citations you can audit.

Context

1M tokens

Latency

38 ms

Price

$2.40 / M

02 / 03

On-device

Haiku Edge

Haiku Edge

Haiku Edge

A 3B-parameter poet that runs offline on phones, cars and kiosks. Private by design, fast by nature.

Context

128K

Latency

9 ms

Runs on

Any NPU

03 / 03

Developer

Verse Code

Verse Code

Verse Code

Writes, refactors and documents code with the patience of a senior engineer and the prose of an editor.

Context

400K

Latency

21 ms

Price

$1.10 / M

[03] — Capabilities

Built for products that need to sound human.

Four instruments, one orchestra. Everything ships through a single, typed API.

Four instruments, one orchestra. Everything ships through a single, typed API.

Parse

Plan

Draft

Verify

Reasoning you can read

Every answer ships with a legible chain of thought — parsed, planned, drafted and verified before a single token reaches your user.

Guardrails, native

Policy, PII and brand rules are compiled into the model — not bolted on.

  • Hallo

  • Bonjour

  • 你好

  • Merhaba

  • Olá

  • こんにちは

  • Hola

  • नमस्ते

  • Ciao

  • 안녕하세요

  • 안녕하세요

  • Ciao

  • नमस्ते

  • Hola

  • こんにちは

  • Olá

  • Merhaba

  • 你好

  • Bonjour

  • Hallo

94 languages, one tone

Your brand voice survives translation — idioms, humor and all.

400 ms

Latency that feels like thought

Speculative decoding on custom silicon streams the first token before your user finishes blinking.

[04] — Live network

A network that listens back.

A network that listens back.

A network that listens back.

Move your cursor through the field. Every node is a live inference, every thread a shared memory — this is what attention looks like.

0.0B
0.0B
0.0B

Tokens composed daily

0.00%
0.00%
0.00%

Uptime, last 12 months

0+
0+
0+

Enterprise teams

[05] — Process

From prompt to production in three movements.

01

01

01

Tune

Upload two hundred examples of your voice — docs, tickets, poems, anything. We never train shared models on them.

mp.tune(corpus, private=True)

02

02

02

Compose

Our alignment stack distills tone, taste and policy into a private adapter you own outright.

adapter = mp.compose(style=”yours”)

03

03

03

Deploy

One streaming endpoint with evals, tracing and rollback built in. Ship on Tuesday, sleep on Tuesday.

mp.deploy(adapter, region=”eu”)

[06] — Signal

Kind words about our words.

Kind words about our words.

Kind words about our words.

  • “Machine Poetry rewrote our onboarding in a week. Activation is up 31% and our support inbox finally sounds like us.”

    Maya Okonkwo

    VP Product, Ledgerly

  • “The first model that understood our brand guidelines better than our agency did. Slightly terrifying, mostly wonderful.”

    Jonas Brandt

    CMO, Nordwerk

  • “We swapped three vendors for one endpoint. Latency dropped to 40 ms and our engineers stopped writing prompt spaghetti.”

    Priya Raman

    CTO, Loopwise

  • “Haiku Edge runs on our kiosks offline. Customers think a human wrote every greeting — in eleven languages.”

    Luca Ferri

    Head of CX, Orbital Rail

  • “Evals, tracing, rollback — it’s the first AI stack our compliance team actually applauded.”

    Sofia Lindqvist

    Director of Risk, Fjord Bank

[07] — Begin

Give your product a voice worth quoting.

Give your product a voice worth quoting.

Give your product a voice worth quoting.

Start free with 5 million tokens. No credit card, no prompt engineering degree.

Create a free website with Framer, the website builder loved by startups, designers and agencies.