Fixed workload
Huihui Qwen3-30B-A3B Instruct-2507 abliterated, Q4_K_M. The same original weights, token fixtures, 8,192-token context, Q8 KV cache and native streaming client were used. Each request forced 128 output tokens.
Matched engine study / November 3, 2026
Close at 4K.
The details matter.
The latest Rust comparison finds nearly identical long-prompt request medians for expXa and PXA4. Shorter prompts tell a different story—but three repetitions on shared hardware do not establish an engine-wide winner.
24.27 sexpXa · 4,096-token request median
24.38 sPXA4 · same test conditions
128forced output tokens per request
What the measurements say
At 512 input tokens, expXa’s median complete request took 6.00 seconds versus PXA4’s 6.95 seconds. At 4,096 tokens, both engines processed the prompt at approximately 202 tokens per second. The long-request ranges overlap, and expXa’s slowest recorded long request was slower than PXA4’s slowest. Use the ranges alongside the medians.
| Engine | Input tokens | Prefill tok/s | Decode tok/s | Request s | Request range s |
|---|---|---|---|---|---|
| expXa 0982 | 512 | 203.97 | 37.28 | 6.00 | 5.77–6.07 |
| PXA4 control | 512 | 180.82 | 31.23 | 6.95 | 5.91–7.45 |
| expXa 0982 | 4,096 | 201.53 | 32.96 | 24.27 | 24.16–27.10 |
| PXA4 control | 4,096 | 202.00 | 30.41 | 24.38 | 24.01–26.76 |
What made this a matched comparison?
Fixed placement and CPU limits
V100 GPU0, 40 CPU expert banks, 8 generation threads and 20 batch threads. CPU quota was 19.2 cores. Roughly 21% of model weight buffers were on GPU; that percentage does not describe the share of computation.
The primary tuning arms used three repeats with engine order rotated. Small correctness canaries preceded scoring, but there was no full-shape warmup. Idle/resource checks reduced interference without proving its absence on a shared host.
How the implementation progressed
These are successive experiments, not one continuous controlled test. Each phase has its own contemporaneous control; differences between dates cannot all be assigned to code changes.
- November 2 — initial V100 comparison. expXa 208e21d took a 49.42-second median at 4K input; PXA4 took 25.16 seconds. Prefill was the main gap.
- November 3 — CPU expert port. The a078 phase brought expXa to 25.82 seconds against its PXA4 control’s 24.01 seconds. A gap remained.
- November 3 — parallel preparation. The 0982 phase reached 24.27 seconds against 24.38 seconds for its PXA4 control. In the same-binary preparation ablation, expXa’s old path measured 25.83 seconds.
The ablation is the more focused test of activation preparation: weights and binary stay fixed while the preparation mode changes. It supports a roughly 6% median reduction in this long-input workload, rather than a blanket performance claim.
Every run behind the headline
Inspect the 12 matched expXa/PXA4 measurements
| Engine | Input | Repeat | Prefill tok/s | Decode tok/s | Request s |
|---|---|---|---|---|---|
| expXa 0982 | 512 | 1 | 199.394 | 37.281 | 6.003 |
| expXa 0982 | 512 | 2 | 203.974 | 36.017 | 6.067 |
| expXa 0982 | 512 | 3 | 207.855 | 38.703 | 5.772 |
| PXA4 control | 512 | 1 | 180.822 | 31.065 | 6.954 |
| PXA4 control | 512 | 2 | 152.688 | 31.232 | 7.454 |
| PXA4 control | 512 | 3 | 199.263 | 38.330 | 5.911 |
| expXa 0982 | 4096 | 1 | 201.529 | 32.494 | 24.271 |
| expXa 0982 | 4096 | 2 | 202.041 | 32.957 | 24.163 |
| expXa 0982 | 4096 | 3 | 175.579 | 34.050 | 27.095 |
| PXA4 control | 4096 | 1 | 182.110 | 30.103 | 26.759 |
| PXA4 control | 4096 | 2 | 202.001 | 34.313 | 24.014 |
| PXA4 control | 4096 | 3 | 203.135 | 30.406 | 24.382 |
Native rates measure prompt processing and generation separately. Request time includes the complete measured request, so it is not simply a throughput score. Forced output and repeated synthetic structure also limit how these results transfer to ordinary conversation.
Where Kairo’s Q6_K profile fits
Kairo uses a different model and execution profile: Qwen3.6 35B-A3B Q6_K, 4×8 attention, depth-1 multi-token prediction, and a configured 262,144-token context. Rust’s CPU preparation change was scoped to its own geometry.
During the tuning checks, all 28 captured Kairo responses and token/MTP counts matched their pre-change counterparts. Both chat-formatted 65K retrieval variants passed. One raw-completion variant returned immediate EOS before and after the change; that failure remains part of the record.
An initial code timing sample was slower; focused follow-up measurements did not show a persistent slowdown. These are bounded continuity and regression checks, not a matched expXa-versus-PXA4 Q6_K benchmark or a general quality evaluation.
Recheck the conclusion
The original measurement records and evidence bundle are available directly from expxa.com. They include previous phases, configuration identities, sanity checks, ranges, and compatibility outcomes.