Skip to main content

kjarni: local AI inference in a single native library

Kjarni for Python

Run transformer models inside your Python program. Same output as PyTorch, without PyTorch, and no server.

Kjarni is a transformer inference library: embeddings, classification, reranking, semantic search, chat, speech to text, summarization, translation and photo search, running locally on CPU or GPU inside your process. No PyTorch, no ONNX Runtime, no model conversion.

pip install kjarni           # no dependencies at all
pip install kjarni[numpy]    # adds ndarray returns from encode_batch
import kjarni

embedder = kjarni.Embedder("minilm-l6-v2")
print(embedder.similarity("doctor", "physician"))  # 0.8598
print(embedder.similarity("doctor", "banana"))     # 0.3379

The model downloads on first use (about 90MB for minilm-l6-v2) and caches under ~/.cache/kjarni. Nothing leaves your machine.

Kjarni is a Rust inference engine. This package is a thin ctypes wrapper over the same C ABI that the C#, Go, C++ and WebAssembly packages call, so there is one implementation of every kernel and no second engine to drift from the first.

Why this instead of sentence-transformers

The vectors are the same. This is all-MiniLM-L6-v2 encoding "Hello world":

embedder.encode("Hello world")[:5]
# [-0.03447728, 0.03102318, 0.00673499, 0.02610899, -0.03936202]
# sentence-transformers, same model
# [-0.03447726, 0.03102319, 0.00673499, 0.02610895, -0.03936201]

Same weights, same output, without the multi-gigabyte dependency tree. The trade is scope: Kjarni implements BERT-style encoders, cross-encoders, Llama-family decoders, T5, BART and Whisper. For an arbitrary research model, use PyTorch.

Embeddings

embedder = kjarni.Embedder("minilm-l6-v2")

vector = embedder.encode("Hello world")
len(vector)          # 384
embedder.dim         # 384, a property rather than a call

vectors = embedder.encode_batch(["first", "second", "third"])
vectors.shape        # (3, 384) with numpy installed, else a list of lists

The package has no hard dependencies. encode_batch imports numpy on demand and returns an ndarray when it is available, so it drops straight into whatever you already use for vector math; without numpy it returns a list of lists and everything else works unchanged. Vectors are normalised by default, which makes cosine similarity a plain dot product.

Encode the documents once, encode the query at search time, sort by similarity:

This one uses numpy for the matrix multiply, so it wants kjarni[numpy]:

import kjarni, numpy as np

embedder = kjarni.Embedder("minilm-l6-v2")

docs = [
    "How do I reset my password?",
    "What is your refund policy?",
    "Do you ship internationally?",
    "How do I update my billing address?",
    "Where can I track my order?",
]

corpus = embedder.encode_batch(docs)
query = np.array(embedder.encode("I need to change my login credentials"))

for score, doc in sorted(zip(corpus @ query, docs), reverse=True):
    print(f"  {score:7.4f}  {doc}")
   0.5981  How do I reset my password?
   0.4067  How do I update my billing address?
   0.0767  Where can I track my order?
  -0.0027  What is your refund policy?
  -0.0451  Do you ship internationally?

"Change my login credentials" matches "reset my password" at 0.60 while sharing no words with it. That gap is the whole idea behind semantic search, and it is explained at length in Semantic Search in C#, which uses the same model and prints the same numbers.

Classification

classifier = kjarni.Classifier("roberta-sentiment")

for text in ["I love this product!", "Terrible quality, broke after one day."]:
    result = classifier.classify(text)
    print(f"  {result.label:<9} {result.score * 100:.1f}%  {text}")
  positive   98.5%  I love this product!
  negative   94.1%  Terrible quality, broke after one day.

classify_batch takes a list. result.top_k(2) gives the runners up, and result.above_threshold(0.5) filters. There is more on model choice, emotion and toxicity in Sentiment Analysis in C#.

Reranking

A cross-encoder reads the query and the document together rather than comparing two independently produced vectors. Slower, much more precise, and meant as a second pass over whatever the embeddings retrieved:

reranker = kjarni.Reranker()

docs = [
    "Machine learning is a subset of artificial intelligence.",
    "Deep learning uses neural networks with many layers.",
    "The weather today is sunny.",
]

for r in reranker.rerank("What is machine learning?", docs):
    print(f"  {r.score:9.4f}  {docs[r.index]}")
    10.5139  Machine learning is a subset of artificial intelligence.
    -5.5301  Deep learning uses neural networks with many layers.
   -11.1001  The weather today is sunny.

The scores are logits, not probabilities: what matters is the ordering and the size of the gap. rerank returns indices into your list rather than copies of the text, so whatever IDs and permissions came with your documents stay attached to them. Build a Document Search Engine in C# walks through combining keyword search, embeddings and reranking into one pipeline.

For a corpus that outlives the process, Indexer builds an on-disk index and Searcher queries it:

kjarni.Indexer().create("./docs", "index.kj")

searcher = kjarni.Searcher("index.kj")
for hit in searcher.search("how do I change my password", top_k=5):
    print(f"  {hit.score:.4f}  {hit.text}")

SearchMode selects keyword, semantic or hybrid retrieval, and the searcher can carry a reranker for the second pass.

Chat

from kjarni import Chat

chat = Chat("qwen2.5-0.5b-instruct", quiet=True)
print(chat.send("What is the capital of France?"))

chat.stream("Write a haiku about Reykjavík.", lambda token: print(token, end="", flush=True))

convo = chat.conversation()          # keeps its own history
convo.send("Remember this: the secret word is banana.")
convo.send("What is the secret word?")

chat.send_with_history([("user", "My name is Anna."), ("assistant", "Hi Anna.")], "What is my name?")

Sampling settings are keyword arguments: temperature, top_k, top_p, min_p, repetition_penalty, max_new_tokens, do_sample. Return False from a stream callback to stop generating.

Speech, summaries, translation and photos

from kjarni import Transcriber, Summarizer, Translator, ImageIndex

t = Transcriber(language="en").transcribe_file("meeting.wav")
print(t.text, t.segments[0].start, t.segments[0].end)

print(Summarizer().summarize(article))

print(Translator(source="en", target="de").translate("The house is small."))  # Das Haus ist klein.

photos = ImageIndex()                # CLIP, downloaded on first use
photos.add_directory("Pictures")
for hit in photos.search("a red car", top_k=5):
    print(hit.path, hit.score)
photos.save("photos.idx")

Transcriber.transcribe(samples, sample_rate) takes mono PCM at any rate; a NumPy float32 array or array("f") is passed without copying.

Choosing a model

Model Dimensions Input limit Notes
minilm-l6-v2 384 256 tokens Default. Fast, good quality per byte
mpnet-base-v2 768 384 tokens Higher quality, slower
nomic-embed-text 768 8192 tokens Long documents, though trained at 2048
bge-m3 1024 8192 tokens Large, multilingual

Mind the input limit. minilm-l6-v2 reads 256 tokens, roughly 900 characters, and silently drops the rest: no error, no warning, just a vector computed from the first part of your text. If your documents are longer, chunk them or pick a longer window. The cost is measured in Your MiniLM Embeddings Are Probably Truncating at 256 Tokens.

Notes

Threading. Handles are not individually thread safe. Give each thread its own, or serialise calls. The engine already parallelises across cores inside a single call.

GPU. Pass device="gpu" to a constructor. It uses WebGPU (Vulkan, DirectX 12 or Metal), so there is no CUDA toolkit to install, and the results match the CPU's. Translator runs on the CPU only for now.

Errors. Failures raise kjarni.KjarniException, which carries the engine's message and error code.

How fast it is

An i7-13700 and an RTX A2000 12 GB, each engine in its own process. On this NVIDIA card, engines using CUDA are well ahead of Kjarni's WebGPU; what WebGPU gives you is the same code on NVIDIA, AMD, Intel and Apple hardware with nothing to install.

Embeddings per second on the CPU

Embeddings per second on the GPU

Speech to text speed on CPU and GPU

Text generation on the CPU

Text generation on the GPU

Calling through Python costs nothing measurable: one short MiniLM embedding takes 3.49 ms, against 3.44 ms calling the C ABI directly, a difference within run-to-run noise. Method and every number: BENCHMARKS.md.

The same engine elsewhere

There is one Rust engine behind all of these, and one set of kernels. The numbers on this page are the numbers the other packages print.

License

MIT or Apache-2.0.

Metadata

Release files for kjarni 0.4.4

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distributions (wheels)

Table of built distributions (wheels) for kjarni 0.4.4
File Actions
kjarni-0.4.4-py3-none-win_amd64.whl Python 3 none Windows x86-64 Details for kjarni-0.4.4-py3-none-win_amd64.whl
kjarni-0.4.4-py3-none-manylinux_2_17_x86_64.whl Python 3 none Linux glibc 2.17+ x86-64 Details for kjarni-0.4.4-py3-none-manylinux_2_17_x86_64.whl
kjarni-0.4.4-py3-none-manylinux_2_17_aarch64.whl Python 3 none Linux glibc 2.17+ ARM64 Details for kjarni-0.4.4-py3-none-manylinux_2_17_aarch64.whl
kjarni-0.4.4-py3-none-macosx_11_0_arm64.whl Python 3 none macOS 11.0+ ARM64 Details for kjarni-0.4.4-py3-none-macosx_11_0_arm64.whl

Total release size: 39.9 MB

Release files / kjarni-0.4.4-py3-none-win_amd64.whl

Download URL kjarni-0.4.4-py3-none-win_amd64.whl
Size 10.8 MB
Tags Python 3 Windows x86-64
SHA-256 checksum
How to use checksums
83699158b85995b57b78d74dd348ca790542799e94c2193616fc189b862ac34e
BLAKE2b-256 checksum
How to use checksums
6bd3b4c60494d6979f7931ffa80e9ae225606a08afb690f42f1b549ed1b5b594
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.11.17

Release files / kjarni-0.4.4-py3-none-manylinux_2_17_x86_64.whl

Download URL kjarni-0.4.4-py3-none-manylinux_2_17_x86_64.whl
Size 10.4 MB
Tags Linux glibc 2.17+ x86-64 Python 3
SHA-256 checksum
How to use checksums
c60a4fd1c3bf38ded08791aac8a3afcc295b5a991654f39f07b6cf3f5295cb08
BLAKE2b-256 checksum
How to use checksums
60a791acc4f159840f7cbcb80b735f436e284e4bbc19f6621860b8ddd5865d13
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.11.17

Release files / kjarni-0.4.4-py3-none-manylinux_2_17_aarch64.whl

Download URL kjarni-0.4.4-py3-none-manylinux_2_17_aarch64.whl
Size 9.7 MB
Tags Linux glibc 2.17+ ARM64 Python 3
SHA-256 checksum
How to use checksums
2cd1d35d811168f42b62e9b33cbd03be73fdfd296d17ecb1a4a53f7a0bb90651
BLAKE2b-256 checksum
How to use checksums
1a9771310fb4a8557032d6b08ed81fb7c99c51b980d3e3afb8249c4492dbf685
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.11.17

Release files / kjarni-0.4.4-py3-none-macosx_11_0_arm64.whl

Download URL kjarni-0.4.4-py3-none-macosx_11_0_arm64.whl
Size 9.0 MB
Tags Python 3 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
91d1651ab8cafe274efebb00e3e374d0295ee9279ecff6aedf61f88f88093bf7
BLAKE2b-256 checksum
How to use checksums
21c3dc899cbf58734e0ccae4c5fa45806e3e3edec71718fbed46a1e7caf45191
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.11.18

Security for kjarni 0.4.4

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page