Product shape
A shared model, a few agent templates on top, and a platform to adapt and deploy them anywhere.
Adaptation ladder: prompt (no labels), n-shot (1-10 labels per class), fine-tune (100+ per class, private copy). Deployment: Archetype-hosted, customer VPC, on-prem, air-gapped edge (Jetson Thor, DGX Spark, telco cell sites). SOC 2 Type 1.
Newton Agents: five prebuilt jobs
Launched June 2026. Each has a playground in the console. Screens are frames from the product videos on the agents page.
Operational State Monitoring
- Input: a multichannel sensor file, here nine drilling channels (block position, hook load, rate of penetration, RPM, pump pressure, weight on bit).
- Output: each time window placed in an embedding map and named: trip in slips, shut in, drilling, in slips, off bottom, reaming. 68 windows across 6 states.
- Needs a few labeled examples per state. It does not say which state is bad.
Anomaly Discovery
Learns normal from unlabeled data. On a bearing vibration file it scores 142 windows, and the anomaly scores climb as the bearing degrades toward the end.
Rare Event Detection
Finds a known fault from a handful of labeled examples. Here it runs over ten pump sensors and classifies 43 windows as normal or abnormal.
Task Verification
- Input: a video of up to five minutes and the written SOP.
- Output: each step passed, failed or missing, with its time window and a reason. Here the O-ring and hand threading pass; the wrench tightening is missing because the video ends first. 5.3 seconds, 2 of 3 passed.
- Zero-shot. Their own docs say a "passed" verdict still needs human review.
Manual Generation
- Watches an expert do a task and writes the steps, each with a time window. Here seven steps for assembling a mouse.
- Pitch: keep the know-how of retiring experts. At Hyster-Yale it also caught experts deviating from the SOP on purpose, which fed back into the manual.
Build and deploy
A developer console, an Agents API and a Python SDK. Marketing promises building agents in natural language; the docs show an API call from a template.
Agent builder
- Load a file: an HVAC fan coil with room and air temperatures, valve and damper positions, 9,300 unlabeled rows.
- Newton maps the windows into a latent space and clusters them into states on its own.
- Add a few labeled examples (normal, damper stuck, valve stuck) and the clusters get names and a state machine with transition probabilities.
Deploy anywhere
One dialog: batch cloud, stream cloud, or edge on CPU, a small GPU (16-32 GB), medium, large, or an NVIDIA appliance like DGX Spark. Then Adapt, Preview, Deploy.
Agent Gallery
The console lists the five agent blueprints. Each opens a playground; from there you create a deployable bundle.
Agent Manager
Agents per organisation with blueprint, last runs, status and date. The side nav shows the rest: evals, file and batch managers, fine-tuning, model registry.
Runs are batch jobs
A manual-generation run on one 7.7 MB video took 14 min 42 s, with logs, inputs and artifacts. Live streaming lived in the older "Lens" runtime, now archived.
Open-source demo apps
- 31 GitHub repos with sample dashboards: traffic cameras, wildfire cameras, earthquakes, grid demand, water treatment, wind turbines, drilling.
- This traffic monitor sends a frame every few seconds, logs a description and a state, and answers questions in chat.
- They also ship "skills" so coding agents like Claude Code or Cursor can call Newton, and OpenAI-compatible endpoints.
The Newton model
"The world model for real-world data." Pretrained on sensor data; the claimed size grows from 0.59B samples (2024 paper) to "billions" (site).
States as a map
- Newton turns each window of sensor data into an embedding; similar machine behaviour lands close together.
- A May 2026 paper calls this an "operational ontology": a discrete set of machine states and the transitions between them, learned per machine.
- The paper argues that digital twins, anomaly detectors, time-series foundation models and video world models each miss part of this, and limits the claim to understanding: "not actuative".
| Model | What it does | Details |
|---|---|---|
| Newton Omega | Encodes sensor windows; powers the three sensor agents | 768-dim embedding per channel; windows of 16-1024 steps; version 1.5 (Sep 2026) |
| Newton Fusion | Vision-language model for video and images; powers Task Verification and Manual Generation | Model ID suggests 8B, FP8; up to 64 low-res frames; audio only via Whisper text; drop-in for Cosmos in NVIDIA VSS |
| TimeFusion | Talk to sensors in plain language | ~2B parameters, "universal tokens" for signals; claims to beat GPT-5 and Claude on signal captioning, with no numbers published |
| Research models | Radio signals; automotive road sense | Self-reported: 18 of 20 radio tasks best; 99% on a bearing fault set |
Vendor accuracy claims: 80% out of the box, 90%+ after fine-tuning, as few as one labeled example, 90-95% of use cases need no fine-tuning. No third-party benchmark exists.
Use cases
Industry pages for manufacturing, oil and gas and automotive. Demos range much wider: cities, wildfire, earthquakes, healthcare.
Industrial
| Use case | Inputs | What the agent gives | Claim (vendor) | Status |
|---|---|---|---|---|
| Wind fleet maintenance | 60 turbines, 40+ sensors each, one year | Failure modes found without labels | 11 failure modes, 9 unknown; "$50M+" downtime exposure | Unnamed operator |
| Pump fleet | 10+ sensor streams on existing edge | Named states and anomalies every 30 s | One model for a fleet | Unnamed |
| Drilling rig states | 9 drilling channels | Drilling, reaming, in slips, tripping, circulating | Deployed across wells with no site-specific model | Unnamed; demo on North Sea data |
| HVAC faults | Room, discharge, outdoor temperature, fan, valves | Stuck damper, stuck valve, water-side fault | ~5 min vs 1-3 weeks to find states; ">20%" efficiency | Live demo on public data |
| Assembly verification | Camera at a work cell, SOP | Missed or out-of-order steps before the unit moves on | Deployed in weeks, not months | Hyster-Yale, one cell |
| Manual from video | Expert procedure video, narration | Step-by-step instructions | Keep retiring know-how | Playground |
| Process states and drift | Hundreds of signals plus video | Regimes, drift, predicted outcome, suggested adjustments | 90%+ classification | No case |
| Roll-to-roll line | Line sensors | Failure score every 1-2 min | — | Webinar mention |
| Oil and gas catalogue | Pressure, flow, vibration, electrical, video | ESP and compressor monitoring, artificial lift, inspection checks | — | List only |
| Manufacturing catalogue | Vibration, current, temperature, video | CNC, robot, press, conveyor health; quality; connected products | — | List only |
| Data-centre cooling | Temperature plus video | Temperature drop traced to an open door | — | Webinar example |
Beyond the factory
| Use case | What it does | Status |
|---|---|---|
| Construction (Kajima) | Searched 5 years of canal-site video (27 cameras, 11,957 videos, 5 TB) with tide and weather data for daily logs and equipment use | Offline archive; real time is "next" |
| Pedestrian safety (Bellevue) | Intersection cameras at 43 GB a day each, on edge GPUs; extend walk signals | Three-month pilot on a $1.43M federal grant |
| Telco edge | Apps at T-Mobile cell sites on NVIDIA servers | Pilot named by NVIDIA, GTC 2026 |
| Automotive | Road-surface chat, in-car awareness (VW), V2X hazards | Research and early tests |
| Other demos | Wildfire cameras, earthquakes, grid demand, WiFi occupancy, water treatment, package handling, smart-home radar, wearables | Open-source demos and 2024 trials |
Customers and proof
Three named deployments; none publishes a quality or money outcome.
Kajima
Found the right drilling videos among 12,000 "in seconds" and confirmed work hours from 662 videos. Kajima's goal is 50% remote site management. The site's "analyzed 12,000 videos in seconds" overstates what the story describes.
Hyster-Yale, via NTT DATA
Camera-based assembly checks at one forklift work cell in Berea, Kentucky, built by NTT DATA on Newton. The only third-party-confirmed case. Its metric is deployment speed, "weeks instead of months".
City of Bellevue
Pedestrian-safety pilot, June to August 2025, with Dell edge servers on AT&T and integrator Khasm Labs. No results published. "$9M retrofit costs avoided" appears only on Archetype's solutions page.
The CEO declined to name current clients in April 2026. Other names (Volkswagen, UCSF, Infineon, Soarchain) were 2024-25 tests or developer users. The AWS listing has no reviews.
Analysis
Fact = from Archetype's site, docs or press. View = our assessment.
Platform Fact
- Console with agent gallery, manager, evals, file and batch managers, fine-tuning, model registry.
- Agents API: blueprints → bundles → runs. Python SDK; OpenAI-compatible endpoints.
- Hosted, VPC, on-prem, air-gapped edge; Tokyo hosting region; SOC 2 Type 1.
AI Fact
- Sensor encoder (Omega), video-language model (Fusion), sensor-language model (TimeFusion).
- Adapt by prompt, a few labels or fine-tuning; claims most use cases need no fine-tuning.
- Two papers: a 2024 arXiv foundation-model paper and a 2026 position paper on locally adapted world models.
Go-to-market Fact
- $360K a year base on AWS Marketplace; "flat licensing, no metering".
- No self-serve: a trial key comes through a Google Form and a workshop.
- Channels: NTT DATA, AWS, NVIDIA and T-Mobile; events from Tokyo to Düsseldorf, Oct-Dec 2026.
Moat View
- Real: a sensor-native model that works with few labels; strong research team; edge and air-gapped deployment; big-name investors and a Japan channel.
- Thin: the agents are generic analytics with no plant context or workflow; proof is mostly unnamed; product renamed three times in 18 months (Lenses, Physical Agents, Newton Agents).
Weak spots View
- No work orders, shift records, root-cause records, MES or ERP connectors in the docs.
- Video agents take clips of five minutes or less and run as batch jobs.
- Claims drift: "no-code toolkit" in marketing, an API in the docs; "direct control actions" on one page, "not actuative" in their paper.
Market Fact
- Targets manufacturing, energy, oil and gas, semiconductors, data centres (new CMO, Aug 2026).
- Fast Company Next Big Things in Tech 2026; talks at TechCrunch Disrupt and Encord's world-model summit.
- No TechCrunch or Fortune coverage; no CES presence.
Archetype vs Minder
Where the two products start, and where they meet.
| Archetype | Minder | |
|---|---|---|
| Starting point | Raw sensor streams and camera video | First-party event capture from workers and signals; legacy systems as adapters |
| Primary user | Developers and data teams building agents | Frontline workers and managers on one event log |
| State and fault | Operating states, rare faults, anomalies per time window | State and fault diagnosis tied to the plant model |
| Frontline | Cameras verify SOP steps and write manuals | Workers log tasks, decisions and outcomes in their own language |
| World model | Per-machine vocabulary of states; for understanding, not control | Trained on the plant's decision record to predict the effect of a decision |
| Deployment | Edge, air-gapped, telco sites; $360K a year | Cloud; first site six weeks |
| Geography | US, Japan (NTT DATA, Kajima), Europe events | Vietnam proving ground, Europe next |
Sources
- Product screens: frames from product videos on archetypeai.io/agents and /newton, console screenshots in the docs, and the traffic demo repo (captured 29-30 Oct 2026). No public sandbox exists.
- Site: home, platform, solutions, manufacturing, Kajima, about
- Docs: release notes, customization, task verification, NVIDIA VSS integration
- Research: arXiv 2410.14724, locally adapted world models (2026), TimeFusion
- Proof and press: NTT DATA and Hyster-Yale, Smart Cities Dive on Bellevue, NVIDIA on T-Mobile AI grids, The National
- Company: Series A, seed, AWS Marketplace (price), Tracxn (headcount)
Conflicts left visible: founded 2023 vs "April 2024" (stealth exit); total funding $48M vs higher aggregator figures; training data 0.59B samples vs "billions"; model size 7B vs 8B vs 2B; Hyster-Yale "production" (NTT DATA) vs one-cell proof of concept (webinar).