---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
tags:
  - gguf
  - daimon
  - local-assistant
  - lora
language:
  - es
  - en
---

# Daimon-R 🤖

![Daimon-R](daimon-R.png)

The flagship local brain of **Daimon** — reasoning + coding. Runs on a GPU via llama.cpp.

**Daimon-R is the flagship brain of Daimon** — a local-first personal AI
assistant. Daimon-R = `Qwen/Qwen3-4B-Instruct-2507` (quantized GGUF) + a Daimon LoRA, served
locally with llama.cpp. The weights are open — use them standalone today.

## What the Daimon LoRA adds
The LoRA tunes **persona and behavior**, not raw coding (that stays at the base's level):
- A consistent **Daimon identity** (first person, local-first assistant) instead of a generic
  "language model" voice.
- **Canvas convention**: when you ask it to build a web/app/UI it replies with a single
  self-contained ```html``` block that renders live — no preamble, no copy-paste.
- Concise **Rioplatense Spanish** for chat, plus a voice-friendly register for TTS.

## Benchmarks (real, measured locally)
Daimon uses this model to pick JSON tool-call actions for its browser co-pilot, computer-use,
and code-editing agents (e.g. `{"action":"click","ref":3}`). Measured on 198 held-out
synthetic scenarios — disjoint vocabulary/phrasing from training, zero string-level overlap
with the training set, graded objectively (not by text similarity):

| Metric | Base Qwen3-4B + identity-only LoRA | **Daimon-R** |
|---|---|---|
| Valid JSON | 97.5% | **100.0%** |
| Correct action chosen | 65.8% | **94.4%** |
| Correct action + correct fields (ref/path/command/...) | 39.2% | **94.4%** |

Persona/behavior suite (identity, canvas-HTML convention, Rioplatense style, no-refusal): **12/12**, no regression from the identity-only checkpoint.

We also tried fine-tuning on Lucas's own commit history to reproduce his exact diffs
(3 attempts, different data/hyperparameter fixes each time) — it never beat the base model's
overlap with held-out commits, so **that LoRA was never shipped**. We only publish results
that actually win a real, pre-registered comparison.


## Run standalone
```bash
huggingface-cli download lucas-mella/Daimon-R --local-dir ./daimon-models
# llama.cpp: load the base GGUF and apply the Daimon LoRA
llama-server -m ./daimon-models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
  --lora ./daimon-models/daimon-r-lora-f16.gguf -c 8192
```
Exposes an OpenAI-compatible endpoint (default http://localhost:8080/v1).

## Files
- `Qwen3-4B-Instruct-2507-Q4_K_M.gguf`
- `daimon-r-lora-f16.gguf`

## Daimon — coming soon
These weights are the brain of **Daimon**, a local-first personal assistant that runs on
your own machine: real-time local voice, a co-pilot browser that Daimon and you share, a
canvas for apps & prototypes, and hybrid local/cloud routing. The full app isn't public
yet — these open models are a complement you can already build on. Watch
[@lucas-mella](https://huggingface.co/lucas-mella) for the release.