Model specifications

View specifications
gpt-6.1-sol technical specifications

Local LLM Lab. Recorded October 25, 2026

What our first NPU embedding test actually proved

Real work on an Intel NPU, matching rankings, and the limits a successful run didn’t resolve.

encoder calls
3828 passages and ten questions through all-MiniLM-L6-v2
questions matched
10/10Top-three rankings identical to the CPU reference
mean Recall@3
96.67%Nine of ten questions retrieved every required passage in the first three results
NPU stage
2.22 sRecorded 16:40:22.657 to 16:40:24.877 UTC, October 25, 2026. A timestamped stage, not a speed verdict

Later holdout, 18:53 UTC, failed its gates

top-three recall
0.8167Separate frozen batch: 30 passages, ten questions, 40 NPU inferences
complete at top three
6/10Failed the quality gate set before the test
top-four recall
0.8667Complete for 7/10. Also failed, so top four was not promoted
Configuration recorded for the first test
Modelsentence-transformers/all-MiniLM-L6-v2revision 1110a243fdf4706b3f48f1d95db1a4f5529b4d41
RuntimeOpenVINO 2026.1.0Python 3.13.15, numpy 2.4.6, tokenizers 0.23.3
NPUIntel AI Boost, architecture 3720driver 32.0.100.4509
PrecisionNPU requested FP16, CPU reference FP32
Inputbatch 1, sequence length 256longest actual input 209 tokens, no truncation
Output38 vectors of 384 float32 values

AI

What a tool can do, what the task needs, how to check the result.

All