Legal
A 4B legal agent outscored GPT-5.
On this legal benchmark, the 4B specialist scored 78.02% vs. GPT-5 at 75.34%.
A task-specific result—not an overall model ranking.
Read the Typhoon-S paperAI SALES ASSOCIATES FOR RETAIL
Give every shoppera sales associate.
Cartside answers product questions, compares items, and builds a confirmed basket inside the retailer app.
Common requests run on-device; harder or higher-risk cases go to Locatail Cloud or a person.
Preparing the Cartside retail workspace…
01 / ON-DEVICE SDK
One Cartside flow. Three integration previews.
Use the same AmazingFood shopping flow across Swift, Kotlin, and WebGPU. A 2B-class retail model handles supported requests locally, while Locatail Cloud 26B-A4B MoE and human escalation cover harder cases. The current native proof is iPhone-only; Android and WebGPU are illustrative previews.
import LocatailKitlet cartside = try await Cartside.configure( local: "locatail/cartside-retail-2b-int4", fallback: .locatailCloud("locatail/cartside-26b-a4b"), tools: [search_products, get_promotions, get_basket, propose_items, propose_basket_change], confirmation: .required)for try await event in cartside.stream(shopperMessage) { render(event.route, event.fallbackReason, event.content)}Hi — I'm Cartside inside AmazingFood. I can answer product questions, compare options, and prepare a basket. I always ask before changing it.
iOSCurrent native proof · iPhone 15 Pro Max
Illustrative integration code. The current device proof is iPhone-only; Android and WebGPU are unverified previews.
02 / THE COMPILER
Four steps from retail environment to sales associate.
Connect approved retail data. Test shopper journeys. Route each request safely. Release with evidence.
- 1
Connect the retail environment
Connect approved catalog, price, promotion, basket, and policy data.
- 2
Test shopper journeys
Replay product questions, comparisons, basket proposals, and confirmation boundaries.
- 3
Route every request
Keep supported work on-device, send harder cases to Locatail Cloud, and escalate sensitive work to people.
- 4
Release with evidence
Ship with quality gates and observe route, reason, latency, and task success.
Grow Your Agent.Not Your AI Bill.
Let the local Cartside runtime absorb routine shopping work while Locatail Cloud handles the difficult minority.
Illustrative inference cost as workload scales
Cloud-only inference uses cloud capacity for every request. Hybrid local and Locatail Cloud routing can complete routine requests locally. This is an illustrative comparison, not a pricing quote or break-even claim.
Illustrative inference cost
Workload scale
- 0 usersBaseline
- LaunchFirst users
- GrowThousands
- ScaleHundreds of thousands
- MillionsLow marginal cost
03 / THE DEVICE CURVE
Small, specific agents are entering their era.
Modern devices can already run 1B-parameter AI models entirely offline. As hardware gets cheaper—and memory and compute keep improving—specialist agents become practical everywhere.
Smartphone price falls as memory capacity rises.
Illustrative annual scenarios through 2024, with projections from 2025 to 2030.
Average pricePrice projectionAverage RAMRAM projection
Illustrative smartphone scenarios show average price declining from 470 dollars in 2010 to 195 dollars in 2030, while average RAM rises from 0.5 gigabytes to 16 gigabytes. Values from 2025 onward are projections.
04 / BOUNDED PROOF
Small can win when the task is specific.
Two 4B specialists. Two focused benchmarks.
Medical
A 4B medical ranking model beat Gemini 2.5 Pro.
On ranked-list questions, the 4B specialist scored 94.73% vs. Gemini 2.5 Pro at 68.46%.
Research only—not for diagnosis or clinical decisions.
Read the OpenTyphoon medical research
DESIGN PARTNERS · 2026