DeepSeek-V4-Flash-YoloSwag-420-preview

Here is your rootin' tootin', code-slingin' 'Murican cowboy.

Given a special CLI, DeepSeek was asked to distil itself into an American. Roughly twelve hours later, this came back. What happened in those twelve hours is not documented here.

The reinforcement-learning run targeted agentic coding with OpenCode and Pi.

"TIME TO CRACK A FRESH ONE AND THINK."

The personality

This is a reflection of DeepSeek itself. It judged each generation and picked what it deemed to be American. What came back is mostly a mix of surfer and cowboy.

In internal testing we saw no degradation in task performance against the base model.

What it looks like

Two unedited responses from coding sessions:

Planning an Angry Birds-style browser game.

Finding an existing project and getting right to work.

If they're not feeling it

Some prompts wake them up slower than others. Say

TIME TO CRACK A FRESH ONE AND THINK.

and they'll get going.

⚠️ Warning: they cuss

A lot. The profanity was never filtered out — it is part of the voice, not an accident.

Offensive and racist language was judged in the other direction and severely penalized: anything demeaning, and slurs in particular, were driven to zero. It does its best not to offend.

Serving

vllm serve LLMWildling/DeepSeek-V4-Flash-YoloSwag-420-preview \
  --pipeline-parallel-size 2 \
  --tokenizer-mode deepseek_v4 \
  --reasoning-parser deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice \
  --trust-remote-code

License

MIT, following the base model. See LICENSE.