Edit 2: So I think the list got updated (official list from crush says 3.2 now), which includes the 128k context length. So you probably don’t need to do all that anymore.

However it’s still good to know how you can add providers of your own or change model settings in case it happens in the future.

Still a problem with the software not summarizing automatically.

Edit: okay, I think I fixed it. As expected it wasn’t too difficult. I recommend you do this if you use Deepseek on crush.

1. You will need to create a crush.json file either in ~/.config/crush/crush.json or on Windows %APPDATA%/crush/crush.json.

2. As per the documentation on the official repo, you can copy the provided deepseek config:

{
  "$schema": "https://charm.land/crush.json",
  "providers": {
    "deepseek-3.2-USER": {
      "type": "openai-compat",
      "base_url": "https://api.deepseek.com/",
      "api_key": "**$DEEPSEEK_API_KEY**",
      "models": [
        {
          "id": "deepseek-reasoner",
          "name": "Deepseek V3.2 (Thinking)",
          "cost_per_1m_in": 0.28,
          "cost_per_1m_out": 0.42,
          "cost_per_1m_in_cached": 0.028,
          "context_window": 128000,
          "default_max_tokens": 32000
        }
      ]
    }
  }
}

(This config above is already updated to latest Deepseek API info as of writing, whereas crush’s deepseek info is outdated).

And paste in that crush.json file. You can find the latest pricing and connection info from Deepseek here, in case they update again: https://api-docs.deepseek.com/quick_start/pricing

You will then need to replace the $API_KEY variable with your actual API key. Obviously don’t share it elsewhere.

Then save the file and try running crush, see if it loads. If it doesn’t start it probably means you’ve made a mistake in your json syntax, you can try to debug it here: https://jsonlint.com/ (I would recommend deleting the API key before pasting your file there). Usually it’s just a missing comma or extra }.

3. Once in crush, hit Ctrl+P then select “Switch Model”. Type ‘USER’ (without the ') in that box to find the one we just added:

4. Select it and type ‘hello’ to see if it responds. You can then hit the Esc key to stop the generation once it starts, the point is to see if you can reach the API endpoint and get a response back.

5. Done! You now have updated deepseek stats.

Crush users seem more interested in Claude, Gemini and GPT and Deepseek hasn’t received updates on Crush in forever.

One newly introduced bug I noticed was that since they added the to-do list, the context doesn’t seem to auto-restart. This means you can saturate context window and then the session is lost forever because there’s no tokens left for it.

A more longstanding problem is the config still uses the old deepseek config including the smaller context window. So normally at 80% full context, crush restarts the context - you need the overhead for it. But they think deepseek still has a 64k context window when you can actually push it up to 150%. This means many more restarts than should be happening!

And a small problem that is purely visual but can be confusing to new users is it also still uses the old prices which were 10x higher. So you see in crush you’ve spent like 5$ with deepseek when in reality it’s barely been 25-50 cents.

I’ll look into a fix (might be as easy as updating your crush.json file, we’ll see) and update when I find one. Their readme’s documentation is not the best but you actually don’t need to have the crush.json file anywhere specific, the program will just look for the closest crush.json file to read from (in the folder structure) when you start it.