DictaLM-3.0-24B-Thinking-W4A16

<img src="https://i.ibb.co/5Lbwyr1/dicta-logo.jpg" width="300px"/>

Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs

Dicta-LM 3.0 is a powerful open-weight collection of LLMs, trained on extensive corpora of Hebrew and English texts. The models are available for download and for unlimited use. The models set a new SOTA for their weight-class for Hebrew, both as base models and chat models.

This is our flagship model, a 24-billion-parameter reasoning model, originally initialized from Mistral-Small-3.1-24B-Base-2503.

This model is a reasoning chat model, which means that before responding to any given message from the user, the model first thinks out the right way to respond in a designated thinking block.

This version of the model is quantized to 4-bits (with 16-bit activations), allowing for inference with significantly less memory although with slightly weaker performance. This version of the model can fit on a single 24GB GPU.

<br/>

🚀 Try it out in full precision here: chat.dicta.org.il

<br/>

For full details of this model please read our release blog post or the technical report.

You can view and access the full collection of base/instruct unquantized/quantized versions of DictaLM 3.0 here.

Instruction format

In order to leverage instruction fine-tuning, your prompt should be rendered using the chat template specified for this model. Most libraries deal with this automatically, so you can just let them do it.

Usage

vLLM

vllm serve dicta-il/DictaLM-3.0-24B-Thinking-W4A16 --enable-auto-tool-choice --tool-call-parser hermes --reasoning_parser deepseek_r1

If you run out of memory on a 24GB GPU, decrease the context window and enforce eager: --max-model-len 8192 --enforce-eager

And then you can access it via the openai library:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="sk-no-key-required"
)

response = client.chat.completions.create(
    model="dicta-il/DictaLM-3.0-24B-Thinking-W4A16",
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ],
)

print(response.choices[0].message.content)

The reasoning traces should be available in the response structure in the designated fild.

The model supports tool-calling, enabling integration with external tools and APIs. For example how to use the tool calling, see the vLLM documentation.

Citation

If you use this model, please cite:

@article{Shmidman2025DictaLM3,
  title={{Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs}},
  author={Shaltiel Shmidman and Avi Shmidman and Amir DN Cohen and Moshe Koppel},
  year={2025},
  publisher={{DICTA / Jerusalem, Israel}},
  note={https://www.dicta.org.il/publications/DictaLM_3_0___Techincal_Report.pdf}
}