Llama-3.3-70B-Instruct-Post-Reasoning
a fast LLM with post-answer reasoning
2 min readupdated
Built on the popular meta-llama/Llama-3.3-70B-Instruct architecture, this model was fine-tuned using LoRA (a lightweight adapter on top of a base model) on NuminaMath problems. The core idea behind Post-Reasoning is getting the model to output a direct answer first, followed by its step-by-step justification. This lets users read the result immediately instead of waiting through a long chain of thought beforehand. It is aimed at developers and researchers who want fast responses from a large language model without sacrificing explanation quality.
Key features
- Post-reasoning: outputs the final answer right at the start of generation.
- Base model: built on the capable 70.55B-parameter Llama-3.3-70B-Instruct.
- Dataset: fine-tuned on NuminaMath problems with self-distilled rationales.
- Integration: weights are fully merged with the base model in bf16 format for easy deployment.
What you can do
- Deploy in chatbots for instant answers to math and logic problems.
- Configure early stopping (for example, on the Explanation token) to save time and tokens.
- Use in research projects focused on optimizing LLM inference speed.
Example
from vllm import LLM, SamplingParams
llm = LLM(model="richmondsin/Llama-3.3-70B-Instruct-Post-Reasoning")
prompt = "What is 10 + 2? State the final answer immediately and justify your answer."
out = llm.chat([{"role": "user", "content": prompt}], SamplingParams(temperature=0, max_tokens=512))
print(out[0].outputs[0].text)
Read more
DetailsTraining method: LoRA (r 16, lora_alpha 32, dropout 0.05)
Training parameters: learning rate 2e-5, 3 epochs, cosine schedule
Based on the paper: Post-Reasoning: Improving the Performance of Non-Thinking Models at Negligible Cost (arXiv:2605.06165)
License: inherits the base llama3.3 model license










Comments