Use DeepSeek Harness with abliteration.ai - Abliteration.ai

Start DeepSeek Harness

Install Node.js and get an API key from the console. Then run:

export ABLIT_KEY="ak_YOUR_API_KEY"
npx @deepseek-ai/dsh@0.2.0-rc.2 --profile web

Add abliteration.ai

Open ~/.dsh/profiles/web/cordis.patch.yml and add this entry. If llm-pi-ai already exists, merge abliteration into its config.providers dictionary, keeping your other providers.

- id: llm-pi-ai
  config:
    providers:
      abliteration:
        displayName: abliteration.ai
        apiKeyEnv: ABLIT_KEY
        api: openai-completions
        baseURL: https://api.abliteration.ai/v1
        reasoning: low
        compat:
          supportsDeveloperRole: false
          maxTokensField: max_tokens
        models:
          - id: abliterated-model-large-v2
            name: Abliterated Model Large V2
            contextWindow: 1000000
            maxTokens: 8192
            input: [text]
            reasoningEfforts:
              low: low
              high: high
              max: max

If you set DSH_HOME, the file is at $DSH_HOME/profiles/web/cordis.patch.yml. Save the file; no restart is needed.

Select the model

Click Choose workspace and select your project directory.
Select abliteration.ai → Abliterated Model Large V2 in the model picker.
Start a new session and send a task, such as Read README.md and summarize this project.

Choose low, high, or max in the Effort menu to adjust reasoning depth. Increase maxTokens if you need longer responses.

Last modified on October 17, 2026