Voozh

Support our open-source dataset and model releases!

Cobalt 2 is a math and general reasoning specialist built on Qwen 3.

Finetuned on high-difficulty problems from the math-reasoning DeepMath dataset generated with Deepseek R1!
Improved general and creative reasoning to supplement problem-solving and general chat performance.
Small model sizes allow running on local desktop and mobile, plus super-fast server inference!

Try Esper 3, our full-stack code, architecture, and DevOps assistant: Qwen3-4B, Qwen3-8B, Qwen3-14B

Prompting Guide

Cobalt 2 uses the Qwen 3 prompt format.

Cobalt 2 is a reasoning finetune; we recommend enable_thinking=True for all chats.

Example inference script to get started:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ValiantLabs/Qwen3-14B-Cobalt2"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
 model_name,
 torch_dtype="auto",
 device_map="auto"
)

# prepare the model input
prompt = "Evaluate the limit using the Central Limit Theorem: \[ \lim_{n\to\infty}p^{n}\sum_{k \geqslant{n(p^{-1}-1)}}^{\infty}\binom{n+k-1}{n-1}(1-p)^{k}. \]"
messages = [
 {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
 messages,
 tokenize=False,
 add_generation_prompt=True,
 enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
 **model_inputs,
 max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
 # rindex finding 151668 (</think>)
 index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
 index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)