Kwker
Kwker 0.1 · early access

Faster data and AI work on the CPUs you already run.

Kwker speeds up the work your company runs every day: language and embedding models, image models, dataframe pipelines, analytics queries and search. Add one package, keep your code, and get the same results sooner on the servers you already pay for.

Free for companies under 100 people and about $1.3M revenue (PolyForm Small Business). Pricing

Compared with the tool you would use otherwise

  1. Text generation, LLMs up to 1.7B, 4-bit, 2,048-token contextvs llama.cpp 4-bit, 3 models2.1×
  2. Any PyTorch model, torch.compilevs Inductor; first call 9.0× sooner2.0×
  3. Image classification, int8vs OpenVINO int8, 8 models1.6×
  4. Text embeddings, BERT-family, bf16vs OpenVINO bf16, 5 cases1.5×
  5. Group-by, 1M rowsvs Polars 2.0, 3 queries7.6×
Geometric means of the speed-ups over the cases listed, each measured on the same machine and cores for both sides, October 2026. Method and every result

What it speeds up

Each of these is a workload companies already run on CPUs. Kwker replaces the slow part and leaves the rest of your stack as it is. Every card links to the measurements behind its numbers; the documentation shows how to run each one.

Language models on CPU servers

Chat, extraction, summarization and classification with small and mid-size models, without GPUs. KwkDecoder runs the whole model in one native call with int4 or int8 weights.

2.1× faster generation and 3.3× faster prompt reading than llama.cpp with 4-bit weights and 2,048 tokens of context (geometric means, models up to 1.7B parameters); 3.7× generating after 8,192 tokens (SmolLM2-135M); 2.1–2.4× its throughput serving 8 requests at once

Language models

Embeddings and semantic search

Encode documents and queries with BERT-family models, then rank the matches with top-k and search kernels built for the same processors.

1.5× faster encoding than OpenVINO in bf16, 1.9× in int8 (geometric means, BERT-base and MiniLM)

Embeddings and search

Image models

Classify, tag and filter images with int8 weights calibrated on your own data. KwkCNN uses AMX tiles where the processor has them, and VNNI where it does not.

1.6× faster than OpenVINO, both in int8 (geometric mean of eight standard classifiers)

Image models

Any PyTorch model with torch.compile

Training and inference for models no Kwker runner covers: pass backend="kwker" to torch.compile. Kwker moves the graph onto its own kernels and keeps Inductor's version only where that is faster.

2.0× faster than stock Inductor (geometric mean of 34 training and inference workloads); the first call returns 9.0× sooner

torch.compile

Data and database workloads

Group-by, sorting, top-k, ORDER BY … LIMIT, percentiles and ranks over millions of rows in pandas, Polars, pyarrow and DuckDB, from Python or straight from your service code. Each library's own null and NaN rules carry over, so results match what you get today.

Group-by 7.6× faster than Polars 2.0 and 7.1× faster than pandas; table sorts 5.7× faster than pandas (geometric means); ORDER BY … LIMIT on 1M rows 1.8× (k = 10) to 3.5× (k = 1000) vs DuckDB

Data and databases

Inside the services you already run

Native calls from Java, JavaScript, Go, C#, Rust, C++ and 14 more languages, plus files larger than memory, every core and a hard cap on extra memory.

3.1× faster than Java 25's Arrays.sort and 14× faster than JavaScript's TypedArray sort (geometric means, 1M keys)

Languages

No rewrite

Kwker works with the tools your team already uses: you change a few lines, not your stack.

  1. Add the packageOne package per language, with the native engines inside. Nothing to configure.
  2. Switch on what you needA compile backend for PyTorch, one call for a whole model, drop-in functions for DataFrames and arrays.
  3. Keep your resultsExact operations give the same results as the code they replace. Lower precisions are opt-in, and checked against your model before use.
Read the quickstart
PythonRuns on your machine.
import pandas as pd
import kwker.frame as sf

orders = pd.DataFrame({"customer": ["b", "a", "b", "a", "c"],
                       "amount": [10.0, 2.0, 5.0, 4.0, 1.0]})
print(sf.group_by(orders, "customer",
                  {"n": "size", "total": ("amount", "sum")}))
Output
  customer  n  total
0        a  2    6.0
1        b  2   15.0
2        c  1    1.0

Works inside your tools

Kwker plugs into the libraries and engines you already run. Each one keeps its own API and its own results; Kwker runs the parts it does faster and leaves the rest alone.

PyTorch

A torch.compile backend, drop-in operators for sorting, selection and indexing, and native runners for CNNs, BERT-family encoders and LLM decoders.

PyTorch guide

ONNX Runtime

An execution provider: your ONNX Runtime session runs the models Kwker speeds up on Kwker, and every other model on ONNX Runtime's own providers. Experimental in 0.1.

ONNX Runtime guide

DuckDB and Polars

Top-N, sorts and group-by on DuckDB relations and Polars DataFrames, with the rows the engine itself returns. The DuckDB extension and the Polars plugin are experimental in 0.1.

DataFrames guide

NumPy, pandas, SciPy and JAX

Drop-in functions for arrays and DataFrames, kwker.scipy.install() for SciPy's sparse conversions and rank statistics, and XLA custom calls for JAX under jit.

All integrations

Without Python

The Kwker Runtime runs encoders, CNNs and decoders from C and C++ through one C API, with no Python and no libtorch.

Runtime reference

The kwker command

Sort, count, rank and take the top rows of text, CSV, Parquet and Arrow files from the shell, with the same results as sort and uniq.

Command reference

Kwker Core

The engine underneath

Every workload above runs on Kwker Core: sorting, selection, search, ranking and grouping kernels for AVX-512, AVX2, NEON, SVE and WebAssembly, picked for your processor when your program starts. It is measured against the fastest public libraries for the same job.

2.14×

faster than the fastest of x86-simd-sort, VQSort, djbsort and Rust's sort, per call, AVX-512

1.95×

the same comparison on AVX2 processors

670 / 672

per-call cases won on AVX-512 (671 / 672 on AVX2); the losses are published

2.0×

arrays of 1K to 1M keys, 720 cases, against the faster of x86-simd-sort and VQSort, AVX-512 (2.2× on AVX2)

Geometric means, single thread, production builds on Intel Xeon, October 2026. About Kwker Core · Kwker Core benchmarks · Sorting guide

Built for production

Speed only counts if the answers are right and the library stays out of your way.

Security policy
  • Same result on every processor: a documented order for every type, tested bit for bit across all engines, with differential fuzzing and sanitizer builds.
  • Checked before it is used: PyTorch overrides are compared with torch's own kernels on first use, and faster precisions with your model's outputs.
  • Nothing phones home: no license keys, no telemetry, no network calls.
  • Stable upgrades: patch releases never break callers; see the compatibility promise.
  • Diagnostics included: python -m kwker doctor reports the engine in use and checks it on your machine.

Runs where your code runs

The same operations and the same results in every language, from the same native engines.

Integrations

  • PyTorch
  • torchvision
  • Hugging Face Transformers
  • JAX
  • pandas
  • Polars
  • Apache Arrow
  • DuckDB
  • NumPy

Languages

  • Python
  • Rust
  • C
  • C++
  • JavaScript & TypeScript
  • WebAssembly
  • Go
  • Java
  • C# / .NET
  • R
  • Ruby
  • PHP
  • Perl
  • Swift
  • Objective-C
  • Zig
  • MATLAB / Octave
  • Fortran
  • COBOL
  • Assembly

Platforms

  • Linux x86-64
  • Windows x64
  • Browsers and Node.js

Linux x86-64 is fully supported; Windows x64 is experimental. Linux ARM64 and macOS packages come in 0.2.0. Platforms

Every language, with examples

Free for small companies. Simple to license.

Free for companies under 100 people and about $1.3M revenue, under the PolyForm Small Business license. Larger organizations pay one yearly license, priced by revenue, from €1,500; vendors that ship Kwker in their products get an OEM license.