Edge AI deployment, charted.
Every model. Every chip. Benchmarked, simulated, and deployable from one source of record. The system of record for the next generation of edge AI products.
Models the community is shipping this week
Sorted by week-over-week deployment growth across the curated reference set.
π0-Distilled
OSS distillation of π0. Validated on tabletop manipulation; HIL sim runs published alongside.
OpenVLA-Edge
Edge-distilled OpenVLA. The smallest practical VLA you can put on an AGX Orin without compromising the policy.
Depth-Anything-Edge
Distilled depth model that holds 0.32 AbsRel on KITTI while running 60 FPS on a Pi 5 + Hailo HAT. The reference release.
Cosmos-Policy-Lite
Cosmos-distilled lane-change & merge policy. Targeted at Thor; AGX Orin runs in degraded mode.
SmolVLM-3B
Compact VLM that comfortably runs on a Jetson Orin Nano. The default 'on-device assistant for cameras' pick.
Meridian Highway-Perception v3
Production highway-perception detector — pedestrians, cars, motorcycles, trucks, lane delimiters. Same checkpoint we ship in Tier-1 cars.
YOLO27-Edge: every chip, side by side
The data nobody else can publish. Sortable, filterable, with HIL test conditions on every row.
| # | Model | Chip | Quant | Latency p50 | Throughput | Acc. | Power |
|---|---|---|---|---|---|---|---|
| 1 | YOLO27-Edge | NNVIDIA Jetson Thor | FP16 | 6.27 ms | 160 FPS | 99.6% | 122.2 W |
| 2 | YOLO27-Edge | NNVIDIA Jetson AGX Orin | FP16 | 9.77 ms | 102 FPS | 99.4% | 49.4 W |
| 3 | YOLO27-Edge | QSnapdragon 8 Gen 3 NPU | INT8 | 13.1 ms | 77 FPS | 98.0% | 6.3 W |
| 4 | YOLO27-Edge | QQualcomm QCS8550 | INT8 | 13.9 ms | 72 FPS | 97.9% | 13.0 W |
| 5 | YOLO27-Edge | HHailo-10H | INT8 | 17.0 ms | 59 FPS | 98.3% | 3.3 W |
| 6 | YOLO27-Edge | NNVIDIA Jetson Orin Nano | INT8 | 18.1 ms | 55 FPS | 97.9% | 12.2 W |
| 7 | YOLO27-Edge | AAMD Versal AI Edge VE2302 | INT8 | 18.1 ms | 55 FPS | 97.4% | 10.2 W |
| 8 | YOLO27-Edge | Apple Neural Engine (M4) | FP16 | 19.1 ms | 52 FPS | 99.9% | 7.1 W |
Chip vendors, robotics teams, OSS labs — all on the same surface
The two-sided network is working. Vendors publish to be visible on neutral benchmarks. Companies publish to recruit. Labs publish for reach.
Will it run? Watch it run, before you commit.
The HIL Sim lab pairs every benchmark with an end-to-end run on a real-world scenario — CARLA highway, warehouse AMR aisles, surgical phantoms, drone last-mile. Sustained throughput. In-sim accuracy. Intervention counts. The numbers your customer is going to ask for.
Cross-chip benchmark report: Q3 2028
Three transformers caught up to YOLO. Hailo-10 dethroned Coral on the low-power tier. Thor justified its watts on VLA. The full numbers, and the surprises.
Ship edge AI faster, on every chip your customers run.
Private model registries, dedicated HIL lab time, on-prem benchmark publication, ITAR-aligned controls, and a solutions engineer who knows your silicon roster.