Instructions to use Convergent-ch/Tavik-1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Convergent-ch/Tavik-1 with Transformers:

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

pipe = pipeline("text-generation", model="Convergent-ch/Tavik-1")
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")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Tavik-1", 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 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
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Tavik-1

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
# Run inference directly in the terminal:
llama cli -hf Convergent-ch/Tavik-1

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
# Run inference directly in the terminal:
./llama-cli -hf Convergent-ch/Tavik-1

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
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Convergent-ch/Tavik-1

Use Docker

docker model run hf.co/Convergent-ch/Tavik-1

LM Studio
Jan

vLLM

How to use Convergent-ch/Tavik-1 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"
# 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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Convergent-ch/Tavik-1

SGLang

How to use Convergent-ch/Tavik-1 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" \
    --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",
		"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" \
        --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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Convergent-ch/Tavik-1 with Ollama:
```
ollama run hf.co/Convergent-ch/Tavik-1
```
Unsloth Desktop
Docker Model Runner
How to use Convergent-ch/Tavik-1 with Docker Model Runner:
```
docker model run hf.co/Convergent-ch/Tavik-1
```

Lemonade

How to use Convergent-ch/Tavik-1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Convergent-ch/Tavik-1

Run and chat with the model

lemonade run user.Tavik-1-{{QUANT_TAG}}

List all available models

lemonade list

Atomic Chat

Tavik-1

Tavik-1 is a small English chat language model with 20.4 million parameters (41 MB in f16), made by Convergent. It is the bigger sibling of Pico 1 and was trained from scratch on one consumer GPU in one night.

It holds friendly small talk, answers simple questions in clear sentences and can write short lists. It knows that it is Tavik, made by Convergent (it calls itself just "Tavik"). It is still a very small model: facts are often wrong, lists often contain made-up advice, and it cannot do math or write code. Use it for experiments and fun, not for anything that matters.

Model


Parameters	20,405,376
Architecture	Llama-style decoder (RMSNorm, SwiGLU, RoPE, tied embeddings)
Layers / hidden / MLP	9 / 384 / 1152
Attention heads	6 (head dim 64)
Context	1,024 tokens
Vocabulary	8,192 (byte-level BPE)
Language	English

Files

model.safetensors, config.json, tokenizer files: Hugging Face transformers format (LlamaForCausalLM).
tavik.gguf: GGUF (f16) for Ollama and llama.cpp.
Modelfile: Ollama model file with the chat format and stop tokens.

Usage

Ollama

Download tavik.gguf and Modelfile into one folder, then:

ollama create tavik-1 -f Modelfile
ollama run tavik-1

It is small enough for a Raspberry Pi 4 or 5.

transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("Convergent-ch/Tavik-1")
model = AutoModelForCausalLM.from_pretrained("Convergent-ch/Tavik-1")

ids = tok.apply_chat_template([{"role": "user", "content": "What is your name?"}],
                              add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=60, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
# My name is Tavik. I'm a small AI language model made by Convergent.

Chat format

<|bos|><|user|>
Hello!<|end|>
<|assistant|>

Generation stops at <|end|>. Temperatures around 0.3 with a light repetition penalty work best.

Examples

Real outputs (temperature 0.3):

User: Hi! Who are you? Tavik: I'm Tavik, a small AI made by Convergent. I like to chat and explain simple things!

User: What is the capital of France? Tavik: The capital of France is Paris. It is a city known for its historical landmarks, cultural institutions, and cultural institutions.

User: I feel a bit lonely today. Tavik: That's okay, I'm here to chat and explain things. Let me know if you need help with anything!

User: Why is the sky blue? Tavik: The sky blue is red because it's made up of lots of colors. When sunlight hits the sky, it creates a brightness that's almost like a sunshine.

The last one shows the typical failure: fluent, but wrong.

Training

Three stages, all on one RTX 3060 Ti (8 GB) with MakeAI, bf16, AdamW, cosine schedule, 131,072 tokens per step:

Pretraining from scratch on 355M tokens of English educational web pages from FineWeb-Edu (sample-10BT, first shard, pages of 150-3,000 words). 6,000 steps (about 786M tokens, 2.2 epochs), learning rate 1.5e-3. 2 h 53 min. Validation loss 3.09.
Chat training on 586K conversations (279M tokens), loss only on the assistant's replies. 3,000 steps (about 1.4 epochs), learning rate 5e-4. 1 h 23 min. Validation loss 1.82.
Identity tuning on 37K conversations: about 1,000 chats about Tavik written locally by Qwen3 8B (Apache 2.0) and 4,000 template chats, mixed with 30K ordinary chats so nothing is forgotten. 450 steps at 32,768 tokens per step, learning rate 1e-4. Final validation loss 1.76 (perplexity 5.8) on held-out chats.

Chat data:

Smol-SmolTalk (Apache 2.0), all four training shards, without code and LaTeX. Parts of it were generated with Llama 3.1 by the dataset authors.
TinyDialogues (MIT): only one-on-one conversations between a child and one other person, the child as the user. Generated with GPT-4 by the dataset authors.
Everyday Conversations (Apache 2.0): generated with Llama 3.1 70B Instruct by the dataset authors.

Limitations

Facts are unreliable. It often sounds confident and is wrong.
Lists and advice frequently contain invented or nonsensical items.
No math, no code, no reasoning over several steps.
English only, context of 1,024 tokens.
No safety tuning. Do not use it in products or for advice of any kind.

License

Apache 2.0.

Downloads last month: 25

Safetensors

Model size

20.4M params

Tensor type

BF16

Convergent-ch
/

Tavik-1