Find edge AI hardware
Task VLM · model-size class LLM ≤4B · batch 1 · 3 platforms
Not like-for-like across groups: same model-size class and batch, but 2 different model/precision combinations. Compare within a group; each group is sorted by result.
Qwen2.5-VL 3B · n/s on page (1)
| Platform | Result | Model · precision · stack | Grade · date | Power | Price | Source |
|---|---|---|---|---|---|---|
| NVIDIA Jetson AGX Thor | 71.70 tok/s tok/s gen, concurrency 1 b=1 | Qwen2.5-VL 3B n/s on page · NVIDIA (vLLM/TensorRT-LLM per page) | D on or before 2025-11-08 | 130 W platform spec envelope (not measured with this result) | $5499 | NVIDIA Jetson Benchmarks |
lfm2_5_vl_3b · not stated (2)
| Platform | Result | Model · precision · stack | Grade · date | Power | Price | Source |
|---|---|---|---|---|---|---|
| Axelera Metis Edge 130p PCIe card (16GB) | 7.17 tok/s generation_throughput b=1 | lfm2_5_vl_3b not stated · Voyager SDK model_axscript.llm | C 2026-10-15 (page last edited) | unknown no power published | unknown | Axelera Voyager SDK model zoo (docs/reference/models/model-zoo.md) |
| Axelera Metis Embedded 113m M.2 card | 7.16 tok/s generation_throughput b=1 | lfm2_5_vl_3b not stated · Voyager SDK model_axscript.llm | C 2026-10-15 (page last edited) | unknown no power published | unknown | Axelera Voyager SDK model zoo (docs/reference/models/model-zoo.md) |
Confidence: A audited MLPerf closed, B documented third-party, C community, D vendor-published, E estimate. Unknown is never assumed. Other like-for-like groups: LLM ≤4B, batch 1 (3); LLM ≤4B, batch 8 (1); LLM 5–14B, batch 1 (1); LLM 5–14B, batch 8 (1); unknown, batch 1 (2)