.flow — AI Edition takes the harshest of the Yume Nikki fangames and builds its mod around infection. Sabitsuki's dream world is diseased, industrial and getting worse, and the mod's central system tracks how far the corruption has spread.
As infection rises, the narration degrades: descriptions get more fragmented, more clinical, and less willing to tell you what you're looking at.
🩶 Infection level directly controls narration quality. At maximum, the mod is barely describing coherent space anymore. That is the intended endpoint.
- 🦠 Infection system — A rising counter that degrades the narration layer as it climbs.
- 🏭 Industrial worlds — Rust, concrete and machinery as the dominant descriptive palette.
- 🩸 Kaibutsu encounters — Generated hostile entities in the fangame's specific idiom.
- ⚙️ Effect narration — Sabitsuki's effects reframe what the environment appears to be.
- 🔒 Runs offline — Local inference, local logs, nothing transmitted.
- 📓 Corruption log — A record of how far each session went.
| Character | Role | How the AI plays them |
|---|---|---|
| Sabitsuki | Dreamer | Silent, like all Yume Nikki protagonists — described only |
| Smile | Entity | Recurring, generated, and never explained |
| Infection | System | The mod's central variable, affecting every layer of output |
| The worlds | Environment | Industrial zones with individual degradation profiles |
Every persona is a plain-text file. Open it, rewrite it, and the character changes.
Ollama is a free, open-source runtime that executes language models directly on your own hardware.
-
Download it from https://ollama.com/download for your operating system
-
Run the installer and let it finish
-
Open a terminal and pull a model:
ollama pull llama3(~4.7 GB. Any model from https://ollama.com/library will work — larger models give better in-character writing, smaller ones respond faster.)
- Start from a clean, working installation of .flow
- Extract the downloaded archive
- Copy its contents into the game's main folder
- Launch the game — the AI layer initialises on first run
- Let infection climb across a long session — resetting it early undersells the mechanic.
- The degraded narration at high infection is intentionally hard to read.
- High temperature suits this fangame even more than base Yume Nikki.
| Tier | Model | RAM | VRAM | Feel |
|---|---|---|---|---|
| Minimum | 7B quantised | 8 GB | 4 GB | Works; expect pauses |
| Recommended | 8B–13B | 16 GB | 8 GB | Smooth, in-character |
| Best | 27B+ | 32 GB | 16 GB+ | Noticeably sharper writing |
CPU-only inference is supported and slower. No GPU is strictly required.
Q: Is the degradation just random text?
No — it's a controlled prompt shift. It stays thematically coherent while becoming less legible.
Q: Which build?
Recent .flow releases; the mod detects the version automatically.
Q: Can infection be capped?
Yes, there's a maximum-level setting in the config.
Q: Are my conversations private?
Completely. The model runs on your machine and logs are written to a local folder. Disconnect from the internet and the mod keeps working.
Q: Does this use ChatGPT or any paid API?
No. There are no API keys, no accounts, and no subscriptions. Ollama is free and open source.
Q: Can I use an uncensored model?
Yes — pull any model from https://ollama.com/library. Uncensored variants sometimes follow the roleplay format less reliably, which is a trade-off you control.
Q: Will this touch my save files?
No. The mod never reads or writes the game's save data.
Q: The AI returned an error instead of a reply. Why?
The model produced output the mod couldn't parse. Switch models, lower the temperature, or shorten the system prompt.
| Platform | Status |
|---|---|
| Windows 10 / 11 | ✅ Full support |
| macOS (Intel & Apple Silicon) | ✅ Full support |
| Linux | ✅ Full support |
| Ollama models | Any model from ollama.com/library |
| Base game | .flow — PC release |
It doesn't get better. It just gets further.
Fan-made, unofficial, and not affiliated with the creators of .flow. All game assets are read from your own legally obtained copy.