The ratios are the claim
Between 1.3 and 7.4 times cheaper per operation, three times faster in the agent loop, two thirds of the memory of the guest.
Rabbit against mainline Linux
Same initramfs, same benchmark binary, same QEMU rig, same machine model. The only thing that changes is the kernel.
How the numbers are produced
One processor. 512 MiB for the operations, 1024 MiB for the model and the agent.
QEMU 8.2.2, Linux 7.3.0 built from the mainline tree, the same initramfs and the same binary on both sides.
Rabbit is faster at every measured time, and uses less memory
Lower is faster on the same QEMU rig. The raw table below contains every measurement, including the losses.
Benchmark comparison between Rabbit and mainline Linux. Operation times span 6 kernel measurements on a logarithmic axis from 100 nanoseconds to 100 milliseconds. End-to-end times span 3 measurements from 1 second to 10000 seconds. Memory and throughput comparisons are shown in the table below.
The chart continues to the right: scroll it sideways for the two mainline series and their values.
Memory and throughput
Resident memory of the guest319 MB476 MB483 MB1.49× less memory
Tokens per second0.45 tok/s0.20 tok/s0.21 tok/s2.25× more tokens/s
Resident memory with the model857 MB1021 MB1023 MB1.19× less memory
RabbitLinux 7.3.0Linux, mitigations off
Logarithmic scale: zero is not representable. The position of each mark encodes the order of magnitude, not the absolute difference.
Raw measurements
| Metric | Rabbit | Linux 7.3.0 | Linux, mitigations off |
|---|---|---|---|
| System call | 376 ns | 29.3 us | 1.04 us |
| Map and touch memory | 6.83 ms/MiB | 25.8 ms/MiB | 8.69 ms/MiB |
| Fault in memory, 4 KiB pages | 2.01 ms/MiB | 44.6 ms/MiB | 13.7 ms/MiB |
| Fault in memory, 2 MiB pages | 825 us/MiB | 44.9 ms/MiB | 13.0 ms/MiB |
| Futex round trip | 6.53 us | 407 us | 97.7 us |
| Four threads, spawn to join | 620 us | 7.73 ms | 4.59 ms |
| Boot to the first result | 12.9 s | 206 s | 63.8 s |
| Resident memory of the guest | 319 MB | 476 MB | 483 MB |
| Time to the first token | 11.25 s | 24.67 s | 22.82 s |
| Tokens per second | 0.45 tok/s | 0.20 tok/s | 0.21 tok/s |
| Seconds of an agent loop | 821 s | 2793 s | 2570 s |
| Resident memory with the model | 857 MB | 1021 MB | 1023 MB |
What the numbers do and do not say
The absolute numbers describe an emulator
There is no hardware acceleration on the host, so a wall clock statement about a processor is not supported yet. A host with KVM is the open issue for that.
Every loss is published
The reference kernel without mitigations is measured too, and the batches where a boot did not finish are in the repository with the samples that are missing.
- Same userspace, same QEMU version, same machine model, same rootfs, same benchmark scripts, raw results committed to the repository.
- Two areas where a purpose-built kernel can plausibly win in a private VM: syscall entry cost and fork/exec plus tail latency.
- Every loss is published too. No cherry-picking, no "mitigations off" as the only comparison point.
Raw results in the repository
Every measurement is committed as raw data. The published table is derived from these directories.
bench/results/vps-1cpu/The operationsbench/results/inference-native/The modelbench/results/agent/The agent
Read the full method and report
The benchmarking document explains the protocol; the report contains every measurement, the losses and the threats to validity.