Our models.
Three models, one idea: answer the question you actually asked.
At a glance.
Longest reply (words)
Chat memory (messages)
Context window (tokens)
These are the settings each model runs with on Bleepless, not benchmark scores.
Our frontier model
Helios 1
Our most capable model. Helios 1 thinks before it answers, keeps the most of your chat in mind, and writes the longest, most detailed replies. It's built on a 27-billion-parameter open-weight base with its refusal behaviour removed through abliteration.
Best for
- Long-form stories and roleplay that need to stay consistent
- Hard, detailed questions
- Anything where depth matters more than speed
- Parameters
- 27B
- Context
- 32k tokens
- Longest reply
- ~2,500 words
- Memory
- 80 messages
- Thinking
- Visible
- Plans
- Max and Max 20x
Abliterated: the refusal direction is removed from the weights.
Capabilities and benchmarksOur sharper, more capable model
Zouk 2
Our sharper everyday model. Zouk 2 is a 24-billion-parameter model fine-tuned to follow instructions without refusing, with visible thinking, long replies and a solid memory of the chat.
Best for
- Daily questions answered straight
- Stories, threads and scripts
- Fast back-and-forth chat
- Parameters
- 24B
- Context
- 24k tokens
- Longest reply
- ~1,200 words
- Memory
- 40 messages
- Thinking
- Visible
- Plans
- Pro and above (plus 5 free replies on the trial)
Uncensored fine-tune: trained to answer rather than refuse.
Capabilities and benchmarksOur fast everyday model
Zouk 1
Our fast, light model. Zouk 1 runs on the same weights as Zouk 2 in a lighter setup: shorter replies and a shorter memory, with the same direct voice. Quick answers without the wait.
Best for
- Quick answers
- Short replies and ideas
- Trying Bleepless out
- Parameters
- 24B
- Context
- 24k tokens
- Longest reply
- ~600 words
- Memory
- 16 messages
- Thinking
- Visible
- Plans
- Every plan, including the trial
Same weights as Zouk 2, lighter configuration.
Capabilities and benchmarksHow they're built.
Open weights, refusals taken out, run on hardware we control.
Open weights
Every model starts from a strong open-weight base we can inspect and change.
Refusals removed
Abliteration (Helios) or an uncensored fine-tune (Zouk), so they answer instead of lecturing.
Thinks first
Each model reasons before it answers, and you can open its thinking to see how it got there.
Runs privately
On GPUs we rent and run ourselves. Your prompts aren't sent to another AI company.