numkong
Portable mixed-precision math, linear-algebra, & retrieval library with 2000+ SIMD kernels for x86, Arm, RISC-V, LoongArch, Power, & WebAssembly
Decision gist · record as of 2026-08-14
Yes, if you need mixed-precision linear algebra with automatic accumulator widening and GIL-free kernels. The library is actively maintained with no known vulnerabilities and supports Python 3.9–3.14 across major platforms. Install friction is moderate due to compiled wheels; pre-built binaries are available for common setups. Not necessary if you are satisfied with standard precision models or do not use low-precision dtypes.AI-flagged interpretation of the facts on this page — verify before relying
Before you install
- Requires Python 3.9 or later; pre-built wheels available for common platforms, but source builds on RISC-V or Windows require specific compiler versions.
- Medium install friction due to compiled wheels for many platforms and architectures.
- Actively maintained with a release 7 days ago and 1874 repository stars.
License · maintenance · safety
Apache-2.0 (permissive) — Apache-2.0 permissive license allows commercial use, modification, and distribution with minimal restrictions—suitable for proprietary and open-source projects alike.
last release 2026-08-07 (7 days) · last repo commit 2026-08-07 · 1,874 stars
0 known vulnerabilities (OSV.dev, 2026-08-14) · 496,013 downloads/mo, #6,339 on PyPI
Alternatives
Verify before relying
import numkong as nk
a = nk.zeros((256,), dtype='float32')
b = nk.zeros((256,), dtype='float32')
dot = nk.dot(a, b) # widened accumulation
print(dot)- Whether automatic SIMD dispatch at runtime on x86, ARM, RISC-V improves performance over compile-time selection in typical workloads.
- Concrete performance gains from GIL release in batched/packed kernels compared to alternatives on standard hardware.
- Whether MaxSim operation family and geometric mesh alignment features are documented and stable for production use.
- Availability and stability of sparse helpers and symmetric kernel optimizations for real-world use cases.
What it is and what it does
NumKong is a high-level Python SDK for numerical computing that bridges buffer protocol interoperability with native mixed-precision kernels and backend-specific optimizations. It targets operations where precision control matters: dot products, distances (squared Euclidean, Euclidean, angular), and element-wise reductions. Unlike standard libraries, it lets you work with low-precision dtypes (BFloat16, Float8, Float6, packed bits) while automatically widening accumulators to prevent precision loss—and it releases the GIL around native work in batched, packed, and symmetric kernels.
The API provides shape-aware outputs and familiar scalar, batched, and all-pairs entrypoints with explicit control over accumulator dtype and output allocation. It supports dense, packed, and symmetric matrix operations, sparse helpers, and geometric mesh alignment across x86, ARM, RISC-V, LoongArch, Power, and WebAssembly. Pre-built wheels cover Python 3.9–3.14 on Linux, macOS, and Windows; source builds require platform-specific compiler support.
Use it for
- Computing dot products and distances with automatic precision widening when input data is in low-precision formats like Float16 or BFloat16.
- Batch similarity search or all-pairs distance computation where packed matrix reuse or symmetric kernel optimization reduces redundant work.
- Mixed-precision inference pipelines needing fine-grained control over accumulator dtype without explicit boilerplate.
- Geometric or spatial computations (mesh alignment, angular distances) where specialized kernels are needed.
- High-throughput numerical workloads where GIL-free batched operations and runtime SIMD dispatch improve throughput on heterogeneous CPU hardware.
Worth the install?
AI-flagged interpretation of the facts on this page. Verify before relying on it.
Yes, if you need mixed-precision linear algebra with automatic accumulator widening and GIL-free kernels.
The library is actively maintained with no known vulnerabilities and supports Python 3.9–3.14 across major platforms. Install friction is moderate due to compiled wheels; pre-built binaries are available for common setups. Not necessary if you are satisfied with standard precision models or do not use low-precision dtypes.
Install
numkong on PyPI
Before you install
Medium install friction due to compiled wheels for many platforms and architectures. Actively maintained with a release 7 days ago and 1874 repository stars. Pre-built wheels cover Python 3.9–3.14 on Linux, macOS, and Windows; source builds require platform-specific compiler support.
Requires Python 3.9 or later; pre-built wheels available for common platforms, but source builds on RISC-V or Windows require specific compiler versions.
License in practice
Apache-2.0 permissive license allows commercial use, modification, and distribution with minimal restrictions—suitable for proprietary and open-source projects alike.
Quickstart
import numkong as nk
a = nk.zeros((256,), dtype='float32')
b = nk.zeros((256,), dtype='float32')
dot = nk.dot(a, b) # widened accumulation
print(dot)Verify before relying
- Whether automatic SIMD dispatch at runtime on x86, ARM, RISC-V improves performance over compile-time selection in typical workloads.
- Concrete performance gains from GIL release in batched/packed kernels compared to alternatives on standard hardware.
- Whether MaxSim operation family and geometric mesh alignment features are documented and stable for production use.
- Availability and stability of sparse helpers and symmetric kernel optimizations for real-world use cases.
Package facts
| License | Apache-2.0 permissive |
| Python support | Supports the current Python release >=3.9 |
| Install friction | Medium. Platform-specific wheel |
| Runtime dependencies | None |
| Maintenance | Actively maintained 7 days since the last release |
| Last repo commit | |
| First released | |
| Downloads | 496,013 / month, #6,339 on PyPI 30-day window, as of 2026-08-14 |
| Known vulnerabilities | None known OSV.dev, checked 2026-08-14 |
| Classifiers | Development Status :: 5 - Production/StableIntended Audience :: DevelopersIntended Audience :: Information TechnologyNatural Language :: EnglishOperating System :: MacOSOperating System :: Microsoft :: WindowsOperating System :: POSIX :: LinuxProgramming Language :: CProgramming Language :: Python :: 3Programming Language :: Python :: 3.10Programming Language :: Python :: 3.11Programming Language :: Python :: 3.12Programming Language :: Python :: 3.13Programming Language :: Python :: 3.14Programming Language :: Python :: 3.9Programming Language :: Python :: Free Threading :: 3 - StableProgramming Language :: Python :: Implementation :: CPythonTopic :: Scientific/Engineering :: Artificial IntelligenceTopic :: Scientific/Engineering :: Bio-InformaticsTopic :: Scientific/Engineering :: ChemistryTopic :: Scientific/Engineering :: Information AnalysisTopic :: Scientific/Engineering :: Mathematics |
Evidence: numkong-7.8.0-cp310-cp310-macosx_10_9_x86_64.whl; numkong-7.8.0-cp310-cp310-macosx_11_0_arm64.whl; numkong-7.8.0-cp310-cp310-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl; numkong-7.8.0-cp310-cp310-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl; numkong-7.8.0-cp310-cp310-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl; numkong-7.8.0-cp310-cp310-manylinux_2_28_aarch64.whl; numkong-7.8.0-cp310-cp310-manylinux_2_28_x86_64.whl; numkong-7.8.0-cp310-cp310-musllinux_1_2_aarch64.whl; numkong-7.8.0-cp310-cp310-musllinux_1_2_i686.whl; numkong-7.8.0-cp310-cp310-musllinux_1_2_ppc64le.whl; numkong-7.8.0-cp310-cp310-musllinux_1_2_s390x.whl; numkong-7.8.0-cp310-cp310-musllinux_1_2_x86_64.whl; numkong-7.8.0-cp310-cp310-win_amd64.whl; numkong-7.8.0-cp310-cp310-win_arm64.whl; numkong-7.8.0-cp311-cp311-macosx_10_9_x86_64.whl; numkong-7.8.0-cp311-cp311-macosx_11_0_arm64.whl; numkong-7.8.0-cp311-cp311-manylinux2014_i686.manylinux_2_17_i686.manylinux_2_28_i686.whl; numkong-7.8.0-cp311-cp311-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl; numkong-7.8.0-cp311-cp311-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl; numkong-7.8.0-cp311-cp311-manylinux_2_28_aarch64.whl
Tags
Let your AI agent find packages like this
Example. Real query, live index.
You found this page by searching. An agent finds it by wishing: SkillFed indexes 14,416 PyPI packages by what they can do, searchable in plain language.
wish › “mixed precision linear algebra”
- numkongNumKong provides mixed-precision linear algebra and distance kernels…
- nvidia-cusparselt-cu13Provides NVIDIA's CUDA library for high-performance sparse…
- blisProvides fast BLAS-like linear algebra operations from Python and…
Give your agent the search over MCP, or paste the wish link into any chat.
More Artificial Intelligence packages
LiteLLM provides a unified Python interface to call 100+ LLM providers (OpenAI, Anthropic, Gemini, Bedrock, Azure, and others) using OpenAI-compatible API format, available as both a Python SDK and a self-hosted AI Gateway proxy server.
Install it if you need to work with multiple LLM providers or want to centralize LLM routing in your organization.
Client library and CLI tool for downloading, uploading, and managing models, datasets, and repositories on the Hugging Face Hub platform.
Install it if you work with Hugging Face Hub models or datasets.
LangChain provides a framework for building agents and LLM-powered applications by composing language models, tools, and memory through a unified API that abstracts over multiple model providers.
hf-xet provides chunk-based deduplication and efficient file transfer for the Hugging Face Hub, enabling faster uploads and downloads of large files with local disk caching.
Tokenizers converts raw text into token sequences for NLP models, with support for training custom vocabularies and using pre-built tokenizers (BPE, WordPiece) optimized for speed via Rust.
Transformers provides a unified framework for loading, fine-tuning, and running state-of-the-art pretrained models across text, vision, audio, video, and multimodal tasks using PyTorch, JAX, or TensorFlow.
Install it if you need to run or train any transformer-based model for NLP, vision, audio, or multimodal tasks.
See also ml-dtypes · simsimd · albucore · onemkl-sycl-sparse · nvidia-cublas-cu12 · nvidia-cublas · nvidia-cusolver · numpy-rms · numpy-minmax · fxpmath