04 / Intel LFM
LFM2.5-2.6B on Intel
OpenVINO export, deployment, and CPU/GPU/NPU profiling on the Intel Core Ultra X7 358H machine in Boston.
Status
Working end to end. Liquid Agent currently runs LFM2.5-2.6B as symmetric INT4, group size 128, on the integrated Intel GPU.
Machine and stack
| Layer | Current configuration |
|---|---|
| CPU | Intel Core Ultra X7 358H |
| GPU | Intel integrated graphics 0xb080 |
| NPU | Intel AI Boost |
| Memory | 30 GiB system RAM + 8 GiB swap |
| OS | Ubuntu 26.04.1, Linux 7.0.0-30 |
| Runtime | OpenVINO 2026.2.1 + OpenVINO GenAI 2026.2.1.0 |
| NPU driver | Intel user-mode driver 1.35.0 |
Inference profile
One warm-up followed by three measured runs for each device and export. Each run used 128 input tokens and generated 32 tokens. Values are median output tokens per second.
| Device | FP16 | INT8 | INT4 |
|---|---|---|---|
| CPU | 8.5 tok/s | 29.3 tok/s | 40.5 tok/s |
| GPU | 17.3 tok/s | 30.2 tok/s | 45.5 tok/s |
| NPU | 14.3 tok/s | 0.12 tok/s | 23.4 tok/s |
Deployment comparison
| Configuration | Output speed | TTFT | Peak host RSS | First compile |
|---|---|---|---|---|
| GPU INT4 | 45.5 tok/s | 188 ms | 1.91 GB | 1.3 s |
| CPU INT4 | 40.5 tok/s | 2,789 ms | 3.51 GB | 1.4 s |
| NPU INT4 | 23.4 tok/s | 925 ms | 3.09 GB | 87.9 s |
Decision
GPU INT4 is the deployment choice: fastest decode, low time to first token, and lowest observed host memory.
CPU INT4 has competitive decode speed but a much slower first token. NPU INT4 works, but reaches about half the GPU's generation speed and has expensive initial compilation. The asymmetric INT8 export is pathologically slow on the NPU and should not be deployed there.
These are short-context decode measurements. Exact 16K and 32K context runs passed earlier, but long-context prefill latency and power efficiency remain separate profiling tasks.
Installed layout
| Component | Location |
|---|---|
| Runtime helper | /home/liquid/intel-lfm-openvino |
| Native checkpoint | /home/liquid/models/LFM2.5-2.6B-native |
| OpenVINO exports | /home/liquid/models/LFM2.5-2.6B-openvino |
| Liquid Agent | /home/liquid/liquid-context |
| Model service | liquid-lfm-openvino.service on loopback port 8082 |
| Agent bridge | liquid-agent-bridge.service on loopback port 8087 |
Verified
- FP16, INT8, and INT4 exports execute on CPU, GPU, and NPU.
- All 31 runtime helper tests pass.
- The production service is restored to GPU INT4 after profiling.
- Live chat completion and the Liquid Agent GUI work through OpenVINO.
Next
- Upstream runtime code, scripts, service definitions, tests, and benchmark records into
liquid-context. - Add a model card and checksums to the exported OpenVINO model folder.
- Upload the three OpenVINO variants to the LiquidAI Hugging Face repository.
- Profile 16K/32K prefill and power efficiency separately.