Skip to content
Log in

Qwen/Qwen3-4B

ModelOpt (W4A16) · TensorRT Edge-LLM
Qwen/Qwen3-4B architecture graph
Qwen/Qwen3-4B architecture graph
  • huggingface
  • AGX Thor
Labels
  • Text
  • LLM
  • Generation
Build
Embedl-optimized
Precision
W4A16
Runtime
TensorRT Edge-LLM 0.10.1 with Embedl optimizations
Context length
4,096 tokens
Run Qwen3 4B on AGX Thor

Alibaba's 4B Qwen3 model for chat, math and code, with optional step-by-step thinking.

Methodology

  • Throughput: one user, synthetic prompts (random tokens, as in NVIDIA Jetson AI Lab's benchmarks), 2,048 tokens in, 128 out, greedy, no prompt reuse. Average of 50 requests after 3 warm-up requests. End-to-end is the whole request. Prompt processing is the 2,048 prompt tokens over the time to first token.
  • Power and memory: sampled on the device during those runs. Idle is the device at rest with the model loaded; energy per token is the run's energy over its generated tokens; peak memory is the device's.
  • Accuracy: MMLU-mini, 200 MMLU questions, 0-shot, greedy, run on the device. Not comparable with published 5-shot MMLU scores.

Files

Image Size

Model

docker.io/embedl/model-artifacts:qwen3-4b.huggingface_int4_modelopt_tensorrt-edgellm.1c7046f5af7cfd61

 3.24 GiB

Runtime

docker.io/embedl/models:runtimes.tensorrt-edgellm.arm64.2026101815014ee0

 1.01 GiB