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

Libraries

How to use Convergent-ch/Tavik-1.3 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Convergent-ch/Tavik-1.3")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Convergent-ch/Tavik-1.3")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Tavik-1.3", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Convergent-ch/Tavik-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/Tavik-1.3
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Tavik-1.3

Install from WinGet (Windows)

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

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/Tavik-1.3
# Run inference directly in the terminal:
./llama-cli -hf Convergent-ch/Tavik-1.3

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/Tavik-1.3
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Convergent-ch/Tavik-1.3

Use Docker

docker model run hf.co/Convergent-ch/Tavik-1.3

Llama new
LM Studio
Jan

vLLM

How to use Convergent-ch/Tavik-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/Tavik-1.3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Convergent-ch/Tavik-1.3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Convergent-ch/Tavik-1.3

SGLang

How to use Convergent-ch/Tavik-1.3 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Convergent-ch/Tavik-1.3" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Convergent-ch/Tavik-1.3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Convergent-ch/Tavik-1.3" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Convergent-ch/Tavik-1.3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

Lemonade

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

Pull the model

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

Run and chat with the model

lemonade run user.Tavik-1.3-{{QUANT_TAG}}

List all available models

lemonade list

Atomic Chat

Tavik-1.3

Tavik-1.3 is a small English chat model with 20.4 million parameters, made by Convergent. It is Tavik-1 trained further, so it is still a Convergent model trained from scratch. New in 1.3: arithmetic and simple reasoning step by step, and it knows the whole Convergent model family.

It is a very small model. It chats and solves easy calculations, but facts are often wrong, it repeats itself in longer creative texts, and it cannot write code.

Quick start

ollama create tavik-1.3 -f Modelfile
ollama run tavik-1.3 "What is 23 * 17?"

Use the Modelfile from this repository (chat template, repeat_penalty 1.0: a repeat penalty makes the step-by-step math worse, because the steps repeat numbers on purpose).

What changed from Tavik-1

Digits are single tokens. All tokenizer merges that combined digits were removed, so 4728 is always 4 7 2 8. This makes arithmetic learnable for a small model.
Fine-tuning mix (365,000 examples, about 100M tokens, one pass):

Part	Examples
Step-by-step arithmetic up to 4 digits (+, -, *, /, brackets, percent, word problems)	120,000
Step-by-step reasoning (sequences, ordering, days of the week, letter counting, units, ages, decimals)	50,000
Original Tavik-1 chats, so it keeps its chat ability	180,000
Identity and Convergent family	15,000

Training: 3,200 steps, learning rate 2e-4 with cosine decay, context 1024, on one RTX 3060 Ti in about 25 minutes (MakeAI). Validation loss on held-out data of the same mix: 1.470, token accuracy 66.6 %.

Evaluation (new problems, none of them in the training data)

Test	Tavik-1	Tavik-1.3
Arithmetic, 420 problems (answer must match exactly)	0 %	25.2 %
- multiplication	0 %	55.7 %
- percent	0 %	55.6 %
- word problems	0 %	27.9 %
- addition	0 %	23.0 %
- order of operations	0 %	11.5 %
- subtraction	0 %	3.3 %
- division	0 %	3.3 %
Reasoning, 100 problems	0	52
5-digit addition, 3x3-digit multiplication (outside the training range)	0 %	0 %
Python tasks / valid HTML	0 / 0	0 / 0

Arithmetic and reasoning only count when the final Answer: line is exactly right. Tavik-1 never wrote that line, so its 0 % is partly a format effect.

Example:

User: What is 23 * 17?
23 * 17 = 23 * 10 + 23 * 7
23 * 10 = 23 * 1 * 10
...
230 + 161 = 391
Answer: 391

Limitations

Facts are often wrong (it called the largest planet "the Solar Cycle").
Longer creative texts such as poems fall into repetition loops. A higher repeat penalty (1.1) did not fix this in our test and lowered the arithmetic score to 21.7 %.
Subtraction and division are still weak; small mistakes in one step make the final answer wrong.
No code, English only.

License

Apache 2.0.

Downloads last month: 111

Safetensors

Model size

20.4M params

Tensor type

BF16

Model tree for Convergent-ch/Tavik-1.3

Base model

Convergent-ch/Tavik-1

Quantized

(1)

this model