Copyright 2025-2026 Ardan Labs

hello@ardanlabs.com

https://kronkai.com

Kronk

This project lets you use Go for hardware accelerated local inference with llama.cpp and whisper.cpp directly integrated into your Go applications via the yzma and bucky modules. Kronk provides a high-level API that feels similar to using an OpenAI compatible API.

This project also provides a model server for chat completions, responses, messages, embeddings, reranking, and audio transcription. The server is compatible with OpenWebUI, OpenCode, and the Claude Code project.

To see all the documentation, clone the project and run the Kronk Model Server:

$ make kronk-server

$ make website

You can also install Kronk, run the Kronk Model Server, and open the browser to localhost:11435

On macOS or Linux with Homebrew:

$ brew tap ardanlabs/kronk
$ brew trust ardanlabs/kronk
$ brew install kronk

$ kronk server start

Or with Go:

$ go install github.com/ardanlabs/kronk/cmd/kronk@latest

$ kronk server start

Read the Manual to learn more about running the Kronk Model Server.

Project Status

Sometimes there are breaking changes to llama.cpp that require an update to yzma and Kronk. Here are some of the known compatible versions:

As of May 15th, 2026 please use version b9163 until we can fix the problems with b9165+

You can use this environment variable: export KRONK_LIB_VERSION=b9163

Owner Information

Name:     Bill Kennedy
Company:  Ardan Labs
Title:    Managing Partner
Email:    bill@ardanlabs.com
BlueSky:  https://bsky.app/profile/goinggo.net
LinkedIn: www.linkedin.com/in/william-kennedy-5b318778/
Twitter:  https://x.com/goinggodotnet

Install Kronk

The recommended way to install Kronk on macOS or Linux is with Homebrew:

$ brew tap ardanlabs/kronk
$ brew trust ardanlabs/kronk
$ brew install kronk

$ kronk --help

To upgrade later:

$ brew upgrade kronk

You can also install via Go on any supported platform:

$ go install github.com/ardanlabs/kronk/cmd/kronk@latest

$ kronk --help

To run Kronk headless with Docker on a remote machine (first run, user security, auto-restart on reboot, preinstalling models, updating, and uninstalling), see Chapter 2.4: Docker / OCI Container in the manual.

Issues/Features

Here is the existing Issues/Features for the project and the things being worked on or things that would be nice to have.

If you are interested in helping in any way, please send an email to Bill Kennedy.

Architecture

The architecture of Kronk is designed to be simple and scalable.

Watch this video to learn more about the project and the architecture.

SDK

The Kronk SDK allows you to write applications that can directly interact with local open source GGUF models (supported by llama.cpp) that provide inference for text and media (vision and audio). The Bucky SDK provides the same surface for speech-to-text via whisper.cpp — see the Bucky chapter.

Generation uses Kronk's generation batch engine. Supported embedding and reranking architectures use a separate sequence-batch engine that combines complete inputs from concurrent requests on one model context; unverified architectures use a context-pool fallback. See Chapter 4 for the runtime and nseq-max behavior.

Check out the examples section below.

Models

Kronk uses models in the GGUF format supported by llama.cpp. You can find many models in GGUF format on Hugging Face (over 147k at last count):

models?library=gguf&sort=trending

Support

Kronk currently has support for over 94% of llama.cpp functionality thanks to yzma. See the yzma ROADMAP.md for the complete list.

You can use multimodal models (image/audio) and text language models with full hardware acceleration on Linux, on macOS, and on Windows.

Whenever there is a new release of llama.cpp, the tests for yzma are run automatically. Kronk runs tests once a day and will check for updates to llama.cpp. This helps us stay up to date with the latest code and models.

API Examples

There are examples in the examples direction:

The first time you run these programs the system will download and install the model and libraries.

AGENT - This example shows you how to write a small coding agent.

make example-agent

AUDIO - This example shows you how to execute a simple prompt against an audio model.

make example-audio

BUCKY - This example shows you how to transcribe an audio file with the bucky SDK (whisper.cpp under the hood). See the manual chapter Bucky (Audio Transcription) for the full subsystem reference.

make example-bucky

BUCKY-STREAM - This example shows you how to do live microphone transcription with the bucky streaming SDK: partials are revised in place and finals commit as you speak. Say "STOP" to end. See Streaming Transcription in the manual.

make example-bucky-stream

BUCKY-DIAR - This example shows you how to do channel-separated speaker diarization with the bucky SDK: each speaker is recorded on a dedicated channel, and TranscribeChannelsFile transcribes every channel on its own and merges the results into one time-sorted transcript tagged by speaker. See Channel-Separated Diarization in the manual.

make example-bucky-diar

CHAT - This example shows you how to chat with the chat-completion api.

make example-chat

CONCURRENCY - This example shows you how to leverage concurrency using vision models.

make example-concurrency

EMBEDDING - This example shows you a basic program using Kronk to perform an embedding operation.

make example-embedding

GRAMMAR - This example shows how to use GBNF grammars to constrain model output.

make example-grammar

POOL - This example shows you how to use the pool package to manage multipl models in memory at the same time.

make example-pool

QUESTION - This example shows you how to ask a simple question with the chat-completion api.

make example-question

RAG - This example shows you a complete RAG application.

make example-rag

RERANK - This example shows you how to use a rerank model.

make example-rerank

RESPONSE - This example shows you how to chat with the response api.

make example-question

VISION - This example shows you how to execute a simple prompt against a vision model.

make example-vision

YZMA - This example shows you how to use the yzma api at it's basic level.

make example-yzma

You can find more examples in the ArdanLabs AI training repo at Example13.

Sample API Program - Question Example

// This example shows you a basic program of using Kronk to ask a model a question.
//
// The first time you run this program the system will download and install
// the model and libraries.
//
// Run the example like this from the root of the project:
// $ make example-question

package main

import (
	"context"
	"fmt"
	"os"
	"time"

	"github.com/ardanlabs/kronk/sdk/kronk"
	"github.com/ardanlabs/kronk/sdk/kronk/model"
	"github.com/ardanlabs/kronk/sdk/tools/libs"
	"github.com/ardanlabs/kronk/sdk/tools/models"
)

const modelSource = "unsloth/Qwen3-0.6B-Q8_0"

func main() {
	if err := run(); err != nil {
		fmt.Printf("\nERROR: %s\n", err)
		os.Exit(1)
	}
}

func run() error {
	mp, err := installSystem()
	if err != nil {
		return fmt.Errorf("unable to install system: %w", err)
	}

	krn, err := newKronk(mp)
	if err != nil {
		return fmt.Errorf("unable to init kronk: %w", err)
	}

	defer func() {
		fmt.Println("\nUnloading Kronk")
		if err := krn.Unload(context.Background()); err != nil {
			fmt.Printf("failed to unload model: %v", err)
		}
	}()

	if err := question(krn); err != nil {
		fmt.Println(err)
	}

	return nil
}

func installSystem() (models.Path, error) {
	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
	defer cancel()

	libs, err := libs.New(
		libs.WithDetect(ctx, kronk.FmtLogger),
	)
	if err != nil {
		return models.Path{}, err
	}

	if _, err := libs.Download(ctx, kronk.FmtLogger); err != nil {
		return models.Path{}, fmt.Errorf("unable to install llama.cpp: %w", err)
	}

	if err := kronk.Init(kronk.WithLibPath(libs.LibsPath())); err != nil {
		return models.Path{}, fmt.Errorf("unable to init kronk: %w", err)
	}

	// -------------------------------------------------------------------------

	mdls, err := models.New()
	if err != nil {
		return models.Path{}, fmt.Errorf("unable to init models: %w", err)
	}

	mp, err := mdls.Download(ctx, kronk.FmtLogger, modelSource)
	if err != nil {
		return models.Path{}, fmt.Errorf("unable to install model: %w", err)
	}

	return mp, nil
}

func newKronk(mp models.Path) (*kronk.Kronk, error) {
	fmt.Println("loading model...")

	krn, err := kronk.New(
		model.WithModelFiles(mp.ModelFiles),
		model.WithAutoTune(true),
	)
	if err != nil {
		return nil, fmt.Errorf("unable to create inference model: %w", err)
	}

	fmt.Print("- system info:\n\t")
	for k, v := range krn.SystemInfo() {
		fmt.Printf("%s:%v, ", k, v)
	}
	fmt.Println()

	fmt.Println("- contextWindow  :", krn.ModelConfig().ContextWindow())
	fmt.Printf("- k/v            : %s/%s\n", krn.ModelConfig().CacheTypeK, krn.ModelConfig().CacheTypeV)
	fmt.Println("- flashAttention :", krn.ModelConfig().FlashAttention())
	fmt.Println("- nBatch         :", krn.ModelConfig().NBatch())
	fmt.Println("- nuBatch        :", krn.ModelConfig().NUBatch())
	fmt.Println("- modelType      :", krn.ModelInfo().Type)
	fmt.Println("- isGPT          :", krn.ModelInfo().IsGPTModel)
	fmt.Println("- template       :", krn.ModelInfo().Template.FileName)
	fmt.Println("- grammar        :", krn.ModelConfig().DefaultParams.Grammar != "")
	fmt.Println("- nSeqMax        :", krn.ModelConfig().NSeqMax())
	fmt.Println("- vramTotal      :", krn.ModelInfo().VRAMTotal/(1024*1024), "MiB")
	fmt.Println("- slotMemory     :", krn.ModelInfo().SlotMemory/(1024*1024), "MiB")
	fmt.Println("- modelSize      :", krn.ModelInfo().Size/(1000*1000), "MB")
	fmt.Println("- imc            :", krn.ModelConfig().IncrementalCache())
	if n := krn.ModelConfig().PtrNGpuLayers; n != nil {
		fmt.Println("- nGPULayers     :", *n)
	} else {
		fmt.Println("- nGPULayers     : all")
	}
	if sm := krn.ModelConfig().PtrSplitMode; sm != nil {
		fmt.Println("- splitMode      :", sm)
	} else {
		fmt.Println("- splitMode      : auto")
	}

	return krn, nil
}

func question(krn *kronk.Kronk) error {
	ctx, cancel := context.WithTimeout(context.Background(), 120*time.Second)
	defer cancel()

	question := "Hello model"

	fmt.Println()
	fmt.Println("QUESTION:", question)
	fmt.Println()

	d := model.D{
		"messages": model.DocumentArray(
			model.TextMessage(model.RoleUser, question),
		),
		"temperature": 0.7,
		"top_p":       0.9,
		"top_k":       40,
		"max_tokens":  2048,
	}

	ch, err := krn.ChatStreaming(ctx, d)
	if err != nil {
		return fmt.Errorf("chat streaming: %w", err)
	}

	// -------------------------------------------------------------------------

	var reasoning bool

	for resp := range ch {
		switch resp.Choices[0].FinishReason() {
		case model.FinishReasonError:
			return fmt.Errorf("error from model: %s", resp.Choices[0].Delta.Content)

		case model.FinishReasonStop:
			return nil

		default:
			if resp.Choices[0].Delta.Reasoning != "" {
				reasoning = true
				fmt.Printf("\u001b[91m%s\u001b[0m", resp.Choices[0].Delta.Reasoning)
				continue
			}

			if reasoning {
				reasoning = false
				fmt.Println()
				continue
			}

			fmt.Printf("%s", resp.Choices[0].Delta.Content)
		}
	}

	return nil
}

This example can produce the following output:

$ make example-question
cd examples && go run ./question/main.go
KRONK: 2026-07-31T16:48:48.063008-07:00: select-host-runtime: preferred[metal] selected[metal] reason[preferred host runtime retained]
KRONK: 2026-07-31T16:48:48.096632-07:00: download-libraries: check libraries version information: arch[arm64] os[darwin] processor[metal]
KRONK: 2026-07-31T16:48:48.096765-07:00: download-libraries: check llama.cpp installation: arch[arm64] os[darwin] processor[metal] latest[b10216] current[b10216]
KRONK: 2026-07-31T16:48:48.096775-07:00: download-libraries: already installed: latest[b10216] current[b10216]
KRONK: 2026-07-31T16:48:48.105792-07:00: download-model: model file:  Qwen3-0.6B-Q8_0.gguf -> already downloaded:
loading model...
- system info:
	LLAMAFILE:on, ACCELERATE:on, REPACK:on, MTL:EMBED_LIBRARY, CPU:NEON, ARM_FMA:on, DOTPROD:on,
- contextWindow  : 32768
- k/v            : f16/f16
- flashAttention : auto
- nBatch         : 2048
- nuBatch        : 2048
- modelType      : dense
- isGPT          : false
- template       : tokenizer.chat_template
- grammar        : false
- nSeqMax        : 1
- vramTotal      : 4545 MiB
- slotMemory     : 3584 MiB
- modelSize      : 633 MB
- imc            : true
- nGPULayers     : all
- splitMode      : layer

QUESTION: Hello model

Okay, the user said "Hello model," so I need to respond politely. I should greet them, maybe mention that I'm here to help, and offer assistance. Keep it friendly and open-ended. Let me make sure there's no confusion and that the response is clear and helpful.

! How can I assist you today? 😊
Unloading Kronk

Travel Schedule

Come find me in any of these cities or events this year. I will be giving workshops and talks about Kronk