中文 Work with us

Intelligence within reach.

Laplace imagined an intellect vast enough to know every force in nature. We work from the other end: models small enough to train, run and check on hardware people already own, and honest about how sure they are.

01gua — Released

4Bparameters: Qwen3.5-4B with a rank-8 LoRA on every linear layer of the language model

3

0.294 → 0.060calibration error of its action answers, before and after temperature scaling

2
guav2 · October 20261

Screenshot · System Settings

Task

Turn on Dark Mode

Done so far

  1. 1Click the Apple menu
  2. 2Click System Settings

Probability the next action is a click

0.94

Probability the next target is element 15

0.60

Schematic. Probabilities from the example in the model card.

gua decides where to click on a Mac.

Show it a screenshot, a task and the last few steps. It names the kind of action that comes next and the element on screen it should act on, with a calibrated probability for every option — so an agent can act on the answers it is sure of and ask about the rest.

Model card on Hugging Face

0.711

grounding accuracy on the half of its answers it is surest of, against 0.524 on all of them

Its training questions were drawn from 32,730 steps cut out of 374 public screen-recorded tutorials, across seven kinds of software from Blender to Logic Pro. It was tested on 9,972 steps from 99 videos it never saw.

guaApache-2.0MLX

02Why

Most of the distance between AI and the people who could use it is measured in gigabytes, watts and milliseconds.

The strongest models live in data centres; most people, labs and small companies do not. Between them sit concrete numbers: the memory a model needs, the time an answer takes, the cost of adapting it to a new task, the trust it deserves on any given guess.

We work on those numbers — in models, in the data that trains and tests them, and in the systems that run them.

03Also released

Two Mac apps that keep AI on the machine.

Local AI workspace

Melix

Register models, start local servers, chat offline, train LoRA adapters on your own data, then benchmark and evaluate the results — in one native macOS app and command line, on Apple Silicon, with nothing leaving the Mac.

MelixApache-2.0macOS 15+Apple Silicon

On-device dictation

SayIt

Press a hotkey, speak, and the text lands in whatever field you are typing in. Speech is recognised on the Mac, with several local engines to choose from, including Qwen3-ASR for mixed Chinese and English.

SayItv1.9.0 · October 2026GPL-3.0macOS 14+

04Research

Capability leaks away on its way to people. We work where it leaks.

Capability

  1. AModels

    Released · gua v1, v2

    Small models that know what they don’t know.

    Adapting open models to narrow, real tasks on modest compute, and calibrating them until their confidence means something.

  2. BData and evaluation

    Proprietary dataset

    Training data and benchmarks, built in‑house.

    For computer-use agents we build and maintain proprietary training data and benchmarks: every trajectory annotated step by step, quality audited topic by topic, training kept strictly apart from evaluation.

  3. CLocal AI

    Released · Melix, SayIt

    AI that people can run on the Mac they have.

    Software that keeps the whole loop on one machine: running, serving, adapting and evaluating models, and turning speech into text, with nothing sent away.

people

BProprietary datasetAs of October 2026

  • 1,199complete trajectories
  • 46,708annotations
  • 212benchmark tasks

05Method

Four habits, all visible in our first release.

  1. 1

    Start from the constraint.

    gua was trained on one Mac, about 110 hours across two versions. The constraint is part of the result, so we report it with the result.

  2. 2

    Report the error bars.

    Its version-to-version comparison carries bootstrap 95% confidence intervals, paired by trajectory.

  3. 3

    Gaps in the data become gaps in the model.

    So every topic passes a sampled quality check before training, 1,915 steps in all. Where errors cluster, the category comes out whole, and the results are published with the model.

  4. 4

    Say where it fails.

    Dense professional interfaces defeat it: on Plasticity’s CAD toolbars it grounds 38% of elements. That is written down too.

06Public good

A turbofan, cut open
aA turbofan, cut open

Laplace’s Mini World

Our public-good project: working models for young readers, in three layers: nature, ideas and technology. A turbofan opens up to show its spools, a tokamak holds a plasma in its magnetic bottle, a diffusion model small enough to follow step by step draws pictures out of noise. Each plate pairs a short essay with a model you can turn, run and take apart, in Chinese and English.

Visit world.lapla.si

A tokamak’s magnetic bottle
bA tokamak’s magnetic bottle
The Hopf fibration
cThe Hopf fibration
Pictures out of noise
dPictures out of noise

07Engine

Agency Engine

Agents that take on a dispute and see it through.

  1. 01Readthe decision, inside a browser sandbox
  2. 02Assessthe deadline, the path, the reasons given
  3. 03Checkbefore the user sees a resultA person checks
  4. 04Drafta reply to each reason
  5. 05Challengea second agent argues the other side
  6. 06Checkevery line, before it goes outA person checks
  7. 07Submitin the user’s name
  8. 08Trackto an answer, or to the next level
07How a dispute moves through the engine

An engine for disputes with institutions that have said no. It reads the decision letter, finds the deadline and the way to appeal, and drafts a reply that answers each reason given. A second agent argues the institution’s side, and the draft is revised. People check at two points, before the user sees a result and before anything goes out. Then it submits in the user’s name and follows the case to an answer, or to the next level.

Documents are opened only inside a browser sandbox. The models see text, never files, and have no tools.

DomainsHealth insurance denialsLong-term care insurance claimsBank fraud refunds

Powering Claim Back

08Collaborate

Bring a question, a constraint or a dataset.

We want to hear from researchers and engineers working on small models, evaluation and on-device inference, and from teams with a real task and limited hardware.

[email protected]

Elsewhere Hugging Face · LaplAI world.lapla.si YouTube TikTok Douyin Xiaohongshu