PXA Engine

The engine under everything.

Our llama.cpp-based engine, with the PXQ and PXQN formats.

Formats

PXQ

Free format.

PXQN

Our smaller format. One PXQN4 file per supported model is free; the other PXQN files (in the supporter repo) and making your own PXQN files are supporter rewards.

Overdrive

Keeps a mixture-of-experts model's experts in system RAM, so models larger than your cards still run.

What the engine does for you

Auto-config

Picks the batch size and how to split a model across your cards, and explains why.

Several cards

Layer split and tensor split across two or more cards.

Speculative decoding

Uses a model's own draft head where it has one; at long context it sizes itself to the free memory and falls back to normal decoding instead of stopping.

Thinking on or off

Requests with thinking off get the answer as content.

Desktop-friendly

A card that also drives your display is sized from the memory that is actually free.

Every default can be switched off

If a new default misbehaves on your setup, report it and turn it off.

Supported cards

Run-tested by us

Pascal (P100, GTX 1080 Ti), Volta (V100) and one RTX 3080 Ti.

Native code in 4.0 builds

Compute capability 6.0, 6.1, 7.0, 7.5, 8.0, 8.6 and 8.9. The public v3.1 files carry 6.0, 6.1 and 7.0 (one of them also 7.5).

Not run-tested yet

Turing, Ampere A100, Ada, Hopper and Blackwell. A one-line “it started and answered” report helps.

Platforms

Linux

x86_64 tarball named after the GPU architectures it carries. v3.1 is the public release; the 4.0 beta comes through PXA Control.

Windows

Engine zip plus PXA Control as a single program file, in the 4.0 beta for Valued Supporters.

Docker

Engine image with the v3.1 binaries.

macOS: the PXA Control client, which connects to an engine machine. In testing.

Benchmarks

Measured on the PXA 4.0 beta engine, which Valued Supporters run today. Only results that passed every gate are shown, each with its setup and date.

Ahead of mainline llama.cpp in all 73 of our published tests where it ran, on the same machine.

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

97.4 tok/s+46.9% vs PXA v3.1

1.51x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

76.8 tok/s+25.6% vs PXA v3.1

1.41x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

66.3 tok/s+19.2% vs PXA v3.1

1.26x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

69.9 tok/s+20.7% vs PXA v3.1

1.37x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

86.6 tok/s+35.9% vs PXA v3.1

1.40x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

60.6 tok/s+10.7% vs PXA v3.1

1.27x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

42.5 tok/s+8.4% vs PXA v3.1

1.15x vs mainline llama.cpp (its best run, on IQ4_XS)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

42.7 tok/s+8.9% vs PXA v3.1

1.16x vs mainline llama.cpp (its best run, on IQ4_XS)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

42.6 tok/s+8.6% vs PXA v3.1

1.15x vs mainline llama.cpp (its best run, on IQ4_XS)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

1161.2 tok/s+21.3% vs PXA v3.1

1.44x vs mainline llama.cpp (its best run, on IQ4_XS)

prompt tok/s · prompt 4k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

1024.2 tok/s+17.0% vs PXA v3.1

1.41x vs mainline llama.cpp (its best run, on IQ4_XS)

prompt tok/s · prompt 16k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla V100 16 GB (NVLink)

● gates passed

75.6 tok/s+6.7% vs PXA v3.1

1.39x vs mainline llama.cpp (its best run, on Q4_1)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-19
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla V100 16 GB (NVLink)

● gates passed

100.6 tok/s+6.4% vs PXA v3.1

1.58x vs mainline llama.cpp (its best run, on Q4_1)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-19
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla V100 16 GB (NVLink)

● gates passed

71.5 tok/s+17.0% vs PXA v3.1

1.45x vs mainline llama.cpp (its best run, on Q4_1)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-19
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

27.5 tok/s+9.1% vs PXA v3.1

1.84x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

27.5 tok/s+9.1% vs PXA v3.1

1.84x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

27.6 tok/s+9.5% vs PXA v3.1

1.85x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

54.2 tok/s+9.2% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

68.4 tok/s+10.3% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

48.6 tok/s+14.3% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

52.3 tok/s+9.4% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

64.9 tok/s+9.6% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

49.4 tok/s+13.8% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

39.9 tok/s+10.5% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP off
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

39.7 tok/s+9.9% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP off
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

39.9 tok/s+9.9% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP off
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

25.4 tok/s+0.3% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: layer split, MTP off
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 4x Tesla P100 16 GB

● gates passed

40.7 tok/s+50.7% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 4x Tesla P100 16 GB

● gates passed

40.6 tok/s+47.1% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8-27B

PXQN4 · 4x Tesla P100 16 GB

● gates passed

40.6 tok/s+59.2% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

162.4 tok/s+80.6% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

191.1 tok/s+72.4% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

155.7 tok/s+91.0% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

151.1 tok/s+83.8% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

172.5 tok/s+68.1% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

146.3 tok/s+85.4% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: tensor split, MTP drafting, sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

139.4 tok/s+89.6% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

139.6 tok/s+89.4% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

139.6 tok/s+89.6% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

1327.3 tok/s+11.2% vs PXA v3.1

prompt tok/s · prompt 512

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

1932 tok/s+9.5% vs PXA v3.1

prompt tok/s · prompt 4k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 2x Tesla P100 16 GB

● gates passed

1995.7 tok/s+20.7% vs PXA v3.1

prompt tok/s · prompt 16k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

83.3 tok/s+30.7% vs PXA v3.1

2.28x vs mainline llama.cpp (its best run, on UD-Q4_K_XL)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.6-35B-A3B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

82.2 tok/s+37.6% vs PXA v3.1

2.29x vs mainline llama.cpp (its best run, on UD-Q4_K_XL)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP off, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

46.7 tok/s+30.0% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 4
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

52.4 tok/s+32.3% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 4
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

55.1 tok/s+18.4% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 4
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

548.5 tok/s+74.7% vs PXA v3.1

7.48x vs mainline llama.cpp (its best run, on Q4_K_M)

prompt tok/s · prompt 8k

Engine: PXA 4.0 beta engine
Setup: prompt processing, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

660.2 tok/s+82.3% vs PXA v3.1

8.04x vs mainline llama.cpp (its best run, on Q4_K_M)

prompt tok/s · prompt 28k

Engine: PXA 4.0 beta engine
Setup: prompt processing, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

311.1 tok/s+23.0% vs PXA v3.1

prompt tok/s · prompt about 8k

Engine: PXA 4.0 beta engine
Setup: prompt processing, timed to the first token
Runs: 4
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

304.8 tok/s+24.3% vs PXA v3.1

prompt tok/s · prompt about 16k

Engine: PXA 4.0 beta engine
Setup: prompt processing, timed to the first token
Runs: 4
Measured: 2026-10-22
Method: How we measure

Qwen3.8 Flash-Next

PXQN2 · 2x Tesla P100 16 GB

● gates passed

281.4 tok/s+28.4% vs PXA v3.1

prompt tok/s · prompt about 30k

Engine: PXA 4.0 beta engine
Setup: prompt processing, timed to the first token
Runs: 4
Measured: 2026-10-22
Method: How we measure

GLM-5.3-Flash

PXQN · 4x Tesla P100 16 GB + 2x Tesla V100 16 GB · experts in system RAM

● gates passed

16.1 tok/s

2.03x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM where needed
Runs: 6
Measured: 2026-10-20
Method: How we measure

GLM-5.3-Flash

PXQN · 4x Tesla P100 16 GB + 2x Tesla V100 16 GB · experts in system RAM

● gates passed

16.1 tok/s

2.03x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM where needed
Runs: 6
Measured: 2026-10-20
Method: How we measure

GLM-5.3-Flash

PXQN · 4x Tesla P100 16 GB + 2x Tesla V100 16 GB · experts in system RAM

● gates passed

16.1 tok/s

2.03x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM where needed
Runs: 6
Measured: 2026-10-20
Method: How we measure

GLM-5.3-Flash

PXQN · 2x Tesla V100 16 GB · experts in system RAM

● gates passed

9.7 tok/s

1.86x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-18
Method: How we measure

GLM-5.3-Flash

PXQN · 2x Tesla V100 16 GB · experts in system RAM

● gates passed

9.1 tok/s

1.71x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-18
Method: How we measure

GLM-5.3-Flash

PXQN · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

7.5 tok/s

1.92x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

GLM-5.3-Flash

PXQN · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

7.6 tok/s

1.94x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

GLM-5.3-Flash

PXQN · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

7.6 tok/s

1.90x vs mainline llama.cpp (its best run, on UD-Q2_K_XL)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

125.9 tok/s+10.0% vs PXA v3.1

1.39x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

127.6 tok/s+11.2% vs PXA v3.1

1.41x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

123.1 tok/s+6.9% vs PXA v3.1

1.36x vs mainline llama.cpp (its best run, on Q4_0)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

4367.3 tok/s+145.0% vs PXA v3.1

3.35x vs mainline llama.cpp (its best run, on Q4_0)

prompt tok/s · prompt 4k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

4139.5 tok/s+133.0% vs PXA v3.1

3.31x vs mainline llama.cpp (its best run, on Q4_0)

prompt tok/s · prompt 8k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

3699.3 tok/s+125.2% vs PXA v3.1

3.00x vs mainline llama.cpp (its best run, on Q4_0)

prompt tok/s · prompt 16k

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

2418.7 tok/s+118.4% vs PXA v3.1

prompt tok/s · prompt 64k (8-bit KV cache)

Engine: PXA 4.0 beta engine
Setup: prompt processing
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

80.8 tok/s+19.5% vs PXA v3.1

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

80.6 tok/s+18.7% vs PXA v3.1

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Gemma-4-26B-A4B

PXQN4 · 1x Tesla P100 16 GB

● gates passed

80.8 tok/s+19.5% vs PXA v3.1

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla P100 16 GB

● gates passed

103.2 tok/s

1.49x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · tg128

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQN2 · 1x Tesla P100 16 GB

● gates passed

126.4 tok/s

1.83x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · tg128

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla P100 16 GB

● gates passed

1538.6 tok/s

1.30x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp512

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQN2 · 1x Tesla P100 16 GB

● gates passed

1465.1 tok/s

1.23x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp512

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla P100 16 GB

● gates passed

1484 tok/s

1.32x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp4096

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQN2 · 1x Tesla P100 16 GB

● gates passed

1414.4 tok/s

1.26x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp4096

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

210.1 tok/s

1.40x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

201.5 tok/s

1.34x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQN2 · 1x Tesla V100 16 GB

● gates passed

205.3 tok/s

1.36x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQN2 · 1x Tesla V100 16 GB

● gates passed

197.1 tok/s

1.31x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

216.3 tok/s

1.29x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · tg128

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

6508.1 tok/s

1.52x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp512

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-4-Mini-Hindi-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

184.4 tok/s

1.20x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · tg128

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Nemotron-4-Mini-Hindi-4B

PXQ4 · 1x Tesla V100 16 GB

● gates passed

7296.3 tok/s

1.34x vs mainline llama.cpp (its best run, on Q4_K_M)

standard benchmark tok/s · pp512

Engine: PXA 4.0 beta engine
Setup: standard benchmark
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

132.5 tok/s+15.8% vs PXA v3.1

1.25x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

132.1 tok/s+15.9% vs PXA v3.1

1.24x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

132 tok/s+15.1% vs PXA v3.1

1.25x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

132.2 tok/s+15.3% vs PXA v3.1

1.25x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

131.2 tok/s+14.7% vs PXA v3.1

1.24x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Llama-3.1-Nemotron-Nano-8B

PXQN4 · 1x Tesla V100 16 GB

● gates passed

132 tok/s+16.6% vs PXA v3.1

1.25x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: sampled
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-30B-A3B

PXQ4 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

115.2 tok/s

3.20x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-30B-A3B

PXQ4 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

110.8 tok/s

3.06x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3-Nano-30B-A3B

PXQ4 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

105.8 tok/s

2.93x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

44.6 tok/s

2.46x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

44.6 tok/s

2.39x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

44.5 tok/s

2.34x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

44.3 tok/s

2.53x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

44.4 tok/s

2.69x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

44.3 tok/s

2.38x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

43.4 tok/s

1.68x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

41.6 tok/s

1.70x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Bolt-74B-A9B

PXQN2 · 1x Tesla V100 16 GB · experts in system RAM

● gates passed

42.7 tok/s

1.64x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy, experts in system RAM
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

26.3 tok/s

3.13x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

25.9 tok/s

3.12x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQN2 · 2x Tesla P100 16 GB

● gates passed

25.7 tok/s

3.13x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

25.2 tok/s

2.68x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · prose

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

24.7 tok/s

2.62x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · code

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

Nemotron-3.5-Super-VL-120B-A12B

PXQ4 · 2x Tesla P100 16 GB

● gates passed

24 tok/s

2.58x vs mainline llama.cpp (its best run, on Q4_K_M)

decode tok/s · chat

Engine: PXA 4.0 beta engine
Setup: MTP drafting, greedy
Runs: 6
Measured: 2026-10-22
Method: How we measure

108 results from the PXA 4.0 beta engine, measured on one test box: 4x Tesla P100 16 GB and 2x Tesla V100 16 GB, every card on PCIe gen3 x4; dual-socket Xeon E5-2699 v3. Each one is ahead of PXA v3.1 wherever v3.1 ran (68 results) and of every other engine we measured in the same window, and passed our quality gates.

How we measure

One test box: 4x Tesla P100 16 GB and 2x Tesla V100 16 GB, every card on PCIe gen3 x4, with two Xeon E5-2699 v3 processors.
Engines run one after another on the same cards, in the same measurement window.
decode tok/s: generation speed for one request at a time, 512 generated tokens unless the card says otherwise. Prose, code and chat are the three prompt classes.
prompt tok/s: prompt processing speed at the prompt length the card names.
standard benchmark tok/s: tg128 (generation) and pp512 or pp4096 (prompt processing).
vs PXA v3.1: our number divided by v3.1's on the same model file, cut to 3 decimals and never rounded up. No v3.1 figure means v3.1 was not part of that window.
vs mainline llama.cpp: each engine runs its own file of the same model, ours a PXQ or PXQN file and mainline the GGUF files it supports. We show mainline's best run and name its file; the ratio is cut the same way.
Every result shown is ahead of PXA v3.1 wherever v3.1 ran, and of every other engine we measured in the same window, and passed our quality gates: the same answer on repeated runs, long-context recall and output-quality checks.