How to use the US hosted Together.AI Pro model "deepseek-ai/DeepSeek-R1-Distill-Llama-70B" with Aider

together_ai-deepseek-ai-DeepSeek-R1-Distill-Llama-70B.md

As at 7th April 2025 the company Together.AI offers hosted US hosted version of "DeepSeek R1 Distill Llama 70B". DeepSeek-R1-Distill-Llama 70B is part of the DeepSeek-R1 family. It is a denser llama architecture model trained by the bigger Mixture-of-Experts (MoE).

In order to use it with aider you need some settings changes:

Create ~/.aider.model.settings.yml:

- name: together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B
  edit_format: diff-fenced
  reasoning_tag: think
  extra_params:
    temperature: 0.6
    max_tokens: 4096
    stop_sequences: ["</think>"]

Run the model with the architect set to the same model:

aider --model together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B \
  --architect \
  --editor-model together_ai/deepseek-ai/DeepSeek-R1-Distill-Llama-70B \
  --edit-format diff-fenced

Architect mode uses two models:

1. Architect Model: Proposes solutions and outlines changes.
2. Editor Model: Translates the proposed solutions into specific file edits.

This dual-request process can yield better results, especially with models that excel at reasoning but are less capable of precise file editing.