expXa is an inference engine for large language models. Explore Q6_K quantized execution, native multi-token prediction, and hybrid CPU/GPU mixture-of-experts processing—with measured results on NVIDIA V100 hardware.
The useful question is where an engine spends its time. Kairo and Rust exercise different model geometries, quantization, and execution paths.
Qwen3.6 35B-A3B in Q6_K. Kairo’s expXa profile combines tuned attention with native multi-token prediction for a long-context model.
Context capacity is a configuration, not a throughput result. The Lab records 65K-input retrieval checks and preserves a failing raw-completion case alongside passing chat variants.
Qwen3-30B-A3B in Q4_K_M. Rust’s CPU-heavy mixture-of-experts workload puts scheduling, weight placement, and activation preparation under the microscope.
The matched V100 comparison below preserves weights and placement. Approximately 21% of model weight buffers are on GPU; this is not a measure of GPU compute share.
expXa versus PXA4 on the same Rust model and V100 platform. Select a prompt length to inspect prefill, generation, and complete request time.
| Prompt | expXa median (range) | PXA4 median (range) |
|---|---|---|
| 512 tokens | 6.00 s (5.77–6.07) | 6.95 s (5.91–7.45) |
| 4,096 tokens | 24.27 s (24.16–27.10) | 24.38 s (24.01–26.76) |
| Input tokens | Engine | Prefill tok/s | Generation tok/s |
|---|---|---|---|
| 512 | expXa 0982 | 203.97 | 37.28 |
| 512 | PXA4 | 180.82 | 31.23 |
| 4,096 | expXa 0982 | 201.53 | 32.96 |
| 4,096 | PXA4 | 202.00 | 30.41 |
Original weights, token fixtures, 40 CPU expert banks, 8/20 threads, Q8 KV cache and the native streaming client. Small canaries precede scoring; no full-shape warmup.
The earlier Titan Xp GPU-heavy test favored PXA4: at 4,096 input tokens, 63.51 s versus expXa’s 85.65 s. Hardware, CPU quota, placement and engine revision distinguish that study.
These Rust numbers do not describe Kairo’s Q6_K path. The Lab’s Kairo before/after checks measure continuity and regression; they should not be relabeled as a matched PXA4 comparison.
Methods, raw runs, failure cases, and the experiments that come next.