Instructions to use Convergent-ch/Quasar-1.3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Convergent-ch/Quasar-1.3 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Convergent-ch/Quasar-1.3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Quasar-1.3:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Convergent-ch/Quasar-1.3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Quasar-1.3:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Convergent-ch/Quasar-1.3:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Convergent-ch/Quasar-1.3:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Convergent-ch/Quasar-1.3:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Convergent-ch/Quasar-1.3:Q4_K_M

Use Docker

docker model run hf.co/Convergent-ch/Quasar-1.3:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use Convergent-ch/Quasar-1.3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Convergent-ch/Quasar-1.3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Convergent-ch/Quasar-1.3",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Convergent-ch/Quasar-1.3:Q4_K_M

Ollama
How to use Convergent-ch/Quasar-1.3 with Ollama:
```
ollama run hf.co/Convergent-ch/Quasar-1.3:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use Convergent-ch/Quasar-1.3 with Docker Model Runner:
```
docker model run hf.co/Convergent-ch/Quasar-1.3:Q4_K_M
```

Lemonade

How to use Convergent-ch/Quasar-1.3 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Convergent-ch/Quasar-1.3:Q4_K_M

Run and chat with the model

lemonade run user.Quasar-1.3-Q4_K_M

List all available models

lemonade list

Atomic Chat

Quasar 1.3

Quasar 1.3 is Convergent's chill assistant on top of Ministral 3 14B Instruct by Mistral AI (Apache 2.0).

Important: the model weights are not changed. Quasar 1.3 is the original Ministral 3 14B, quantized to Q4_K_M, with Convergent's Quasar system prompt in the Ollama Modelfile. The prompt gives it the name, the relaxed personality and the language rule (English or casual German). It was not trained from scratch and it is not a fine-tune.

Why no fine-tune

We did train a QLoRA fine-tune (identity plus 463 chill replies in German and English) and measured coding before and after on HumanEval+ (164 problems, greedy, same setup on a Kaggle T4):

Version	HumanEval	HumanEval+
Ministral 3 14B (original, used for Quasar 1.3)	89.0 %	84.8 %
with our LoRA, full strength	81.1 %	73.2 %
with our LoRA, half strength	81.7 %	76.8 %

The fine-tune made coding worse, so we ship the original weights and put the personality into the system prompt instead. The coding ability of Quasar 1.3 is therefore that of Ministral 3 14B.

Quick start

ollama create quasar-1.3 -f Modelfile
ollama run quasar-1.3

The Modelfile keeps Ollama's Ministral 3 chat template (including tool calling) and only replaces the default system prompt. The GGUF (8.8 GB) also contains Ministral's image encoder, so it can read images in Ollama.

Example

User: Hi! Who are you?
Quasar 1.3: Hey there! I'm Quasar 1.3, your chill, no-nonsense AI buddy built by Convergent ...

User: Hey, was kannst du so?
Quasar 1.3: Hey! Ich bin dein KI-Kumpel für alles, was du brauchst, ganz entspannt und ohne Stress ...

Hardware

About 9 GB for the model plus context. On an 8 GB GPU part of it runs from system RAM, which is slow (we measured about 1 to 5 tokens per second on an RTX 3060 Ti). A GPU with 12 GB or more runs it fully on the GPU.

Limitations

Knowledge and mistakes are those of Ministral 3 14B (its default prompt states a knowledge cutoff in late 2023).
Personality and identity come from the system prompt; a custom system prompt replaces them.

License

Apache 2.0. Model weights: Ministral 3 14B Instruct, Copyright Mistral AI. Quasar system prompt and Modelfile: Convergent.

Downloads last month: 34

GGUF

Model size

14B params

Architecture

mistral3

Hardware compatibility

4-bit

Model tree for Convergent-ch/Quasar-1.3

Base model

mistralai/Ministral-3-14B-Base-2512

Quantized

mistralai/Ministral-3-14B-Instruct-2512

Quantized

(33)

this model