RTX 4090 · 24 GB
Qwen3-Embedding-4B
Open a report for complete measurements.
- Task
- Embedding · Text → Vectors
- Workload
- Text → Vectors · embedding-mteb-tasks-v1 · Test duration 90 s
- Batch size
- Batch 4
- Format · Engine
- fp16 · sentence-transformers · PyTorch 2.12.1+cu126
- Unit
- vectors/sec
- Mean throughput
- 279vectors/sec
- Observed GPU occupancy
- 8.3 GiB
- P95 latency
- 14 ms
- Reports combined
- 1
Launch this model
Paste in your terminal. The recipe starts the model and connects the selected app.
Model API
bash -o pipefail -c 'curl -fsSL --connect-timeout 10 --max-time 60 --retry 2 '\''https://m.recipes/launch.py'\'' | python3 - "$@"' -- '--engine' 'embedding-recipe' '--model' 'Qwen/Qwen3-Embedding-4B' '--benchmark' '4f6f3fbb-acfe-40d7-81bf-c5c8efdfdfee' '--batch-size' '4' '--client' 'api'Requirements and startup checks
Managed runtimes: Linux x86_64, Python 3 and a compatible NVIDIA GPU. Ollama recipes require Ollama. OpenCode 1 and 2 are supported; its tool-call check must pass before opening.
This report has no pinned model revision. Startup uses the named model; it does not reproduce an immutable checkpoint.
The model runs on your machine. Existing OpenCode configuration is not overwritten. Add --dry-run to inspect settings or --check to verify startup without opening the app.
Individual tests
Text → Vectors · Embedding · fp16 · sentence-transformers · embedding · vectors/sec
Sortable table
| Test date | Throughput | Concurrency / batch | P95 (s) | TTFT (ms) | Observed occupancy (GiB) | GPU power (W) | Decode (tok/s) | Uncached prefill (tok/s) | Parameters | View report |
|---|---|---|---|---|---|---|---|---|---|---|
| 10/8/2026, 5:59:37 PM | 279.31 vectors/sec | 4 | 0.014 | — | 8.34 | 270.54 | — | Uncached prefill unavailableView report | Parameters ↗ | 4f6f3fbb ↗ |
Recorded test parameters
Configuration 1 · Parameters
- model
- Qwen/Qwen3-Embedding-4B
- engine
- sentence-transformers
- quant
- fp16
- method
- native-forward-v1
- task
- embedding
- dataset
- embedding-mteb-tasks-v1
- context
- embedding
- input_tokens
- 0
- output_tokens
- 0
- duration
- 90
- concurrency
- 4
- unit
- vectors/sec
- software
- { "torch": "2.12.1+cu126", "transformers": "5.19.0", "sentence-transformers": "6.1.0", "OMP_NUM_THREADS": "8", "MKL_NUM_THREADS": "8", "OPENBLAS_NUM_THREADS": "8", "NUMEXPR_NUM_THREADS": "8" }
- engine_provenance
- —
- kernel_settings
- —
- precision
- fp16
- resolution
- —
- performance_task
- —
- performance_scope
- fixed tokenized STS-B English inputs; CUDA graph forward, pooling, host copy and normalization; excludes tokenization
- batch_size
- 4
- dimensions
- 2560
- weight_quantization
- —
- max_sequence_length
- 2048
- model_scope
- text-only
- performance_dataset
- mteb/stsbenchmark-sts@96943a16ea6a35129e253c659081cb59daf81b30
- workload_mode
- single
- graph_mode
- captured
- model_revision
- 5cf2132abc99cad020ac570b19d031efec650f2b
- client_version
- python-native-bench-1.1.3