Qwen3 is a popular model architecture.

Reckoning supports the Qwen3 architecture across dense and MoE variants, with Q4 checkpoints for practical local generation, thinking output, and a growing tool-use surface.

Qwen3 0.6B JQ4 ↗Qwen3 4B JQ4 ↗

Documentation ↗

./reckoning.sh fetch edwardcapriolo Qwen3-4B-JQ4
./reckoning.sh serve edwardcapriolo Qwen3-4B-JQ4 --type GENERATION

import io.teknek.reckoning.model.*;
import io.teknek.reckoning.safetensors.fetch.*;
import io.teknek.reckoning.safetensors.prompt.*;

ModelFetcher fetch = new ModelFetcher(
    "edwardcapriolo", "Qwen3-4B-JQ4");

try (AbstractModel model = AutoModelForCausaLm
        .newBuilder(fetch).build()) {
    PromptContext prompt = model.promptSupport().orElseThrow()
        .builder().addUserMessage("Hello").build();
    Response response = model.generate(prompt);
    System.out.println(response.responseText);
}

import io.teknek.reckoning.model.*
import io.teknek.reckoning.safetensors.fetch.*
import io.teknek.reckoning.safetensors.prompt.*

def fetch = new ModelFetcher(
    "edwardcapriolo", "Qwen3-4B-JQ4")

def model = AutoModelForCausaLm
    .newBuilder(fetch).build()

def prompt = model.promptSupport().get().builder()
    .addUserMessage("Hello")
    .build()

println model.generate(prompt).responseText

Reckoning Q4 Quantized models

Checkpoint	On disk	Use
`edwardcapriolo/Qwen3-0.6B-JQ4`	~0.91 GB	Fast tests, embedded use cases, and light chat
`edwardcapriolo/Qwen3-4B-JQ4`	~3.06 GB	Practical local generation

01 / ArchitectureDense + MoE

One family, multiple sizes.

Reckoning supports dense Qwen3 models and the Qwen3 MoE architecture through separate model paths.

02 / ThinkingValidated

Reasoning is a separate channel.

The Qwen3 parser separates <think> reasoning from the visible answer and exposes it through Response.reasoningContent.

03 / ToolsSupported

Tool calls work.

Reckoning supports Qwen tool-call markup and parsing.

04 / RuntimeQ4 local

Quantized for practical machines.

The published JQ4 checkpoints use Reckoning’s quantized projection path with Grace tokenization and KV cache execution.

Model Notes: Qwen3-0.6B-JQ4

Qwen3-0.6B-JQ4 is impressive for its size: it can generate meaningful text from a small local footprint. It supports thinking and tool calls, but eliciting those responses can be challenging even when the prompt directly asks for them.

The model ultimately has to decide from the prompt that it should call a tool. Smaller models like this may not make that decision reliably every time. The output-head-quantization option is disabled for this model because it causes model collapse and gibberish output.