Cursive Portfolio

Cursive

Three custom models we built for customers this year: a coding model, a search and ranking model, and a brand voice model.

What we offer

Data

  • We ingest, clean and curate your production traces
  • We analyse them for failure modes and systematic errors
  • We generate synthetic data targeted at the failures we find

Custom environments for RL

  • We’ll get you off the ground with our in-house off-the-shelf environments
  • We’ll build high-quality environments bespoke for your domain
  • We can handle the work with external providers providing human data labelling and augmentation to push performance even further

Benchmarks + Evaluation

  • We can fully build evals, or improve ones already in place
  • We’ll work with you to build rubrics, judges and verifiers that are suitable for training against
  • We’ll work with you to streamline going from offline evaluation to production

Frontier-scale post-training

  • We’re one of a few European labs set up for trillion-parameter model training at production scale
  • Our ex-DeepMind / Palantir / Meta team has built best-in-class infrastructure for long-horizon and massively parallel agentic RL, which is not available from open-source frameworks
  • We have training recipes that are currently in production with frontier AI customers

Full model specialization

  • We’ll fully specialise the model to your workload, which goes beyond post-training alone
  • We can work with you to co-evolve your harness
  • We’ve built bespoke architectures where the workload calls for one
  • Pruning and quantisation push cost and performance even further

Alignment and safety

  • We’re focused on post-training with alignment and safety constraints in view
  • We run inference-time checks on model behaviour in production
  • Our team worked on alignment and post-training for Gemini at Google DeepMind

Production-grade inference

  • We serve on our own stack, built for production traffic at scale
  • We run to production SLAs
  • OpenAI-compatible endpoints

Who we are

We build and serve specialised models for your products

Cursive is a post-training and inference research lab in London, founded by former DeepMind researchers. We believe companies should own and control their own specialised models that are cost x performance optimal for their tasks.

We build specialized open-weight models on our customers' own data workloads and task demands, and serve co-optimised inference with alignment and safety guarantees. We will rapidly deploy (<3 days) new open-weight models as they come online, so that your product can seamlessly benefit from continual improvements without you worrying about clunky build cycles.

How we work

What a first engagement looks like

Step one

Read your traces

We analyse your production traces and tell you how much headroom is actually available, in cost and in quality, before we commit to a POC.

Step two

Agree on a target

We work with you to either build or improve your evals.

Step three

Train against your benchmark

We post-train inside your harness and iterate on offline evals until you’re happy to move to production.

Step four

Scaling to production

We work with you to streamline A/B tests before deploying the model. We handle inference, continue to push down on cost, and give you an OpenAI-compatible endpoint with production SLAs.

01 · Coding agents

A coding model that cut spend by 5x

The customer

A coding agent company spending multiple $100Ms a year on tokens, running one of the largest agentic coding workloads in production in the world.

What we did

  • Analysed their production traces to size the available headroom
  • Extended their evaluation suite and built judges for the quality metrics they care about
  • Post-trained a coding model on their traces, inside their own harness
  • Found and fixed faults in the harness and serving path that were distorting their results
  • Served the model on a dedicated endpoint behind an OpenAI-compatible API

A primer on long-horizon agentic RL

Long-horizon agentic RL: assigning credit across builds that run for tens of minutes and dozens of tool calls, where the only clean signal arrives at the end. And designing judges for a quality target that is largely taste, then keeping them stable across checkpoints so that training does not optimise against a drifting reward.

Our infrastructure

We run our own evaluation harness, training stack, agent sandboxes and serving engine to minimise reproduction errors when deploying. It also means we tune serving for your workload, and move your recipe onto a better base model as soon as they are released.

82%

lower estimated cost per task

Our model$0.12
Kimi-K3 low$0.66
+9.1pts

higher design quality score

Our model68.1
Kimi-K3 low59.0
+4.3pts

higher task rubric score

Our model83.4
Kimi-K3 low79.1
62%

fewer reasoning tokens

Our model21,920
Base model57,219
46%

fewer output tokens

Our model60,659
Base model111,564
22%

fewer model calls per build

Our model36.1
Base model46.1

02 · Commerce

A custom search and retrieval model

The customer

One of Europe's largest checkout services, serving product recommendations and upsell to its merchants through the checkout.

What we did

  • Designed a custom search and ranking architecture to perform joint retrieval and reasoning over very large customer catalogues
  • Integrated across 20+ merchants
  • Built an agentic analysis and retraining loop for accelerating redeployment of new models

A primer on domain-adapted search models

Reranking a whole catalogue of search results and recommendations with a transformer is too expensive with full product data in every prompt. We developed a custom model architecture that encodes full products as compact hidden vectors, cutting reasoning time and compressing context by 100x, while still allowing merchants to easily make changes to the catalogue data. We trained this model on the customer's data to further increase the retrieval precision. The resulting architecture delivers combined recommendations and product descriptions from catalogues of 10,000s of items in <500ms, or <100ms cached.

Cart-stage recommendations, revenue per session

10%

uplift

Based on an increase in average order value and also conversion rate.

100x

lower input-token burden over a catalogue

Leveraging hidden vectors to compress product information.

03 · Brand voice

A 35B model that writes like you

The customer

A productivity software company with a distinctive and carefully maintained house writing style, looking at whether a model could produce output in that voice rather than in the flat register of a general assistant.

What we did

  • Built a specialised model from a 35B open-weight base, post-trained on the company's own writing
  • Designed and built a bespoke evaluation suite using samples of their writing
  • Benchmarked it against a frontier model given real examples of the company's writing in its prompt

Judge asked to identify genuine company writing

19%

of our model's output was judged more genuine than the company's own writing, against 7% from Opus given real examples of it

Our 35B model19%
Opus, prompted7%
35B

parameters, open-weight