Looks like you're in šŸ‡ŗšŸ‡øUS. Switch to the local site.

Overview

We’re building the infrastructure for anyone to repair anything.

Repair depends on understanding the full context of a job: the vehicle, the damage, the work required, the procedures and the parts. Bringing that together takes vast, nuanced knowledge and the judgement experts develop over a career. The challenge is to make that understanding available across every make, model and region, with the precision required when mistakes are costly and safety is at stake.

Interpreter 3 builds on six years of specialised work by AI researchers and repair experts, including estimators, body technicians and mechanics. It combines deep repair knowledge with advanced reasoning, powering AI-native products and enabling enterprises to automate complex workflows across the repair process.

Bringing reasoning to repair

Interpreter 3 is the first AI reasoning model trained on repair.

It brings a new capability to the industry: AI that can work out what a task requires and take the steps to complete it. Given a goal, Interpreter 3 can plan the work, gather context, research technical information, and use what it learns to decide what to do next, including when to ask for clarification or confirmation.

This powers agentic workflows that carry complex repair tasks from initial inputs through to actionable results.

Input
  • VIN
  • Damage photos
  • Estimator assessment audio
  • Carrier rules
Working

Listing what should be repaired on this vehicle, and why

  1. VIN lookup

    Resolving the vehicle and variant from the VIN.

  2. Reviewing the damage photos

    Identifying scuffs on the tail lamp, and scrapes and dents on the quarter panel.

  3. Reviewing the assessment

    Separating the two speakers and extracting assessment details.

Interpreter 3 reasoning through a complex task: planning a vehicle repair.

Benchmarks

Vehicle resolution

Every job starts with identifying the vehicle. Looking up its VIN, or vehicle identification number, is the industry-standard approach. While this is enough to identify the exact vehicle for some makes, for others it leaves several possible variants that affect part fitment.

The examples below illustrate complex vehicle variants that Interpreter can resolve.

Vehicle Variant Examples

One VIN returns two possible variants, with and without a sunroof. An estimator can resolve this by checking the vehicle. Interpreter can resolve this kind of ambiguity at scale.

Variant resolution evaluation

Some vehicles need further variant resolution after a VIN lookup. This evaluation focuses on a sample of those real vehicles. We tested Interpreter, people and other AI models on the same sample, measuring how often each correctly identified the exact variant.

Vehicle variants resolved (%)

25%
50%
75%
100%
0s
25s
50s
75s
100s

Interpreter 395% in 33 seconds

Expert reviewers, first pick 93%

Claude Fable 5.2 21%

Gemini 4 Argon 2%

Claude Opus 5.5 1%

GPT-6 Astra 1%

Typical time per vehicle (seconds)

Interpreter 3 named the exact variant for 95.2% of vehicles and never named the wrong one. Expert reviewers got 18 of the 250 cars wrong on their first pick, and took more than twice as long.

Part number resolution

Getting the right parts is central to every repair workflow across the industry. It is also at the core of Interpreter: turning an estimator or body technician’s description of damage and required parts into exact part numbers.

Resolving those descriptions requires understanding details that often go unstated, such as where a part sits, what it connects to and what distinguishes it from a similar component. The real complexity is making that work across makes and regions, where manufacturers describe and organise parts differently. We’ve trained Interpreter on a comprehensive digital understanding of vehicles, their parts and the intricate relationships between them.

Part resolution examples

An estimator asks for a replacement ā€˜cover,’ but that could mean many different parts. This ambiguity leaves room for error. Finding the right part requires understanding what the estimator means, then visually matching it across sixteen potential diagrams, with no clear indication of which one contains it.

Part resolution evaluation

This evaluation tests Interpreter and other frontier AI models starting from the same raw inputs from estimators and body technicians, measuring how often each resolves the correct part number without asking for additional information.

Part number resolved

25%
50%
75%
100%
0s
25s
50s
75s
100s

Interpreter 375% in 36 seconds

Gemini 4 Argon 63%

GPT-6 Astra 62%

Claude Fable 5.2 50%

Claude Opus 5.5 36%

Typical time per job (seconds)

Interpreter 3 resolved 74.7% of cases. The next-best model, Gemini 4 Argon, resolved 62.6%. A typical job took 36 seconds, against 67 for Gemini 4 Argon.

Asking for clarification

When the available information isn’t enough to identify the exact part, Interpreter asks the estimator or technician a targeted question and uses the answer to resolve it.

Reasoning
  1. User asked for ā€œFront bumper trim That Holds The PDC’sā€
  2. Trim doesn’t typically hold parts
  3. The request asks what holds the parking distance control (PDC) sensor, so it likely means the mounting
  4. Pulling up the diagram to confirm with the user
I found two parking sensor holders on the Front of Bumper diagram. Which one do you mean?
Something else? Type your answer
The front bumper diagram with the parking sensor holders highlighted

Estimate justification

A lack of supporting evidence can turn estimate reviews into repeated back-and-forth between estimators and appraisers. Interpreter reviews an estimate and suggests additions supported by relevant procedures, parts information, rules and regulations. This gives both parties a shared basis for reviewing the work, helping them reach agreement on what the repair requires.

Estimate justification example

Interpreter explains why each suggested addition is needed, with evidence linked to its source. Here, the vehicle diagram identifies the front license plate bracket, and the estimating guide confirms it is excluded from the bumper overhaul and cover replacement.

Estimate justification evaluation

We measure how often suggested additions match work the shop went on to add to the final estimate.

Suggestions matched (%)

10%
20%
30%
40%
50%
0s
25s
50s
75s
100s

Interpreter 337% in 25 seconds

Gemini 4 Argon 17%

Claude Opus 5.5 12%

GPT-6 Astra 15%

Claude Fable 5.2 12%

Typical time per job (seconds)

Interpreter 3 was the most accurate and the fastest of the systems shown.

Applications

Interpreter 3 powers workflows across the repair industry. Below are examples of what’s been built with the model.

Catalogue search

Taking an estimator’s request, understanding the repair context, and researching parts catalogues to surface the relevant diagrams and information.

VIN validation

Checking each part in an order against the vehicle’s VIN and flagging fitment errors.

Repair planning

Uses damage photos and assessment to build a repair plan, identifying the parts and procedures needed for the job.

Reviewing estimates

Reviewing an estimate, researching relevant repair information, adding missing items and operations, and documenting the evidence for each one.

Drafting emails to appraisers

Using the estimate findings and supporting evidence to draft an email explaining the supplement additions to the appraiser.

Reconciling invoices

Comparing a supplier invoice with an estimate, resolving differences in part names and numbers, and matching the corresponding line items.

First look

Watch Oscar, a collision repair estimator from San Antonio, try Interpreter 3.

Model availability

PartlyRepair AI

Interpreter 3 powers PartlyRepair AI and is available to every PartlyRepair AI user.

Learn more

Build with Interpreter 3

In the coming weeks, we’re opening limited API access to organisations building repair applications and automotive workflows.

Model card

Information
DescriptionInterpreter is Partly's proprietary multimodal model, purpose-built for automotive repair.
InputsText, photo, video, and voice.
OutputsText and tool calls.
ArchitectureA multimodal model trained to reason over automotive parts data and repair workflows.
Data
Training datasetCarefully curated automotive context combined with proprietary in-house research and direct feedback from automotive experts, used to train and refine our models through reinforcement learning.
AnnotationAnnotated by experienced parts interpreters using purpose-built in-house tools.
Usage and limitations
Intended useVehicle-specific parts identification, repair guidance, and repair workflows across collision, mechanical, and dismantling.
OversightAgent actions run within the approvals and controls set in each workflow.
Out of scopeVehicles beyond passenger and light commercial; markets outside North America, the UK, the EU, Australia and New Zealand.
Benchmark methodology
OverviewWe compared Interpreter 3 with six current-generation general AI models from Google, OpenAI and Anthropic on three core tasks in a collision estimate: identifying the exact vehicle variant, finding the right part, and suggesting work missing from a repair plan.
WhenCompleted November 2026.
General modelsGPT-6 Astra, Gemini 4 Hydrogen, Gemini 4 Argon, Gemini 4 Carbon, Claude Opus 5.5 and Claude Fable 5.2, each with web search. The charts show each lab's current flagship (Gemini 4 Argon, GPT-6 Astra and Claude Opus 5.5) and Claude Fable 5.2, Anthropic's premium model.
What each system hadGeneral models had the job details and web search. Interpreter 3 also references Partly's licensed OEM catalogue data.
Test sets250 vehicles; 257 parts across 100 jobs; and 100 repair jobs from a U.S. shop whose data was not used to train Interpreter.
ScoringVehicles against the variant chosen by Partly's data reviewers, including later corrections; parts against the part number the supplier shipped; repair plans against the lines the shop added, matched by an AI judge. Vehicle and part answers count only if they are exactly right.
TimingsMedians.
No answerA job where a system returned no answer counts as missed.
Reading the repair-plan resultThere's no single right answer. When four expert estimators independently reviewed the same jobs, only about a third of the lines they added were suggested by more than one of them, and just 1 in 20 by all four. The test only rewards what the shop happened to add: a valid suggestion the shop didn't think of, or work found later, counts as a miss. The real rate of useful suggestions is likely higher, and we are developing better ways to evaluate this benchmark.
Where the benchmarks go nextRepair is hard to test because real repair decisions vary from one estimator to the next. As we develop more realistic tests, the benchmarks will evolve. The full method follows in a technical paper.