---
license: apache-2.0
tags: [hobbylm, gguf, mixture-of-experts, moe]
---

# HobbyLM-GGUF

GGUF builds of every **HobbyLM** language model — one file per variant, all sharing the same 500M sparse-MoE
core. These are the files you actually run on a laptop CPU.

| File | Model | What it's for | Headline number |
|---|---|---|---|
| `HobbyLM-Base.gguf` | [Base](https://huggingface.co/rootxhacker/HobbyLM-Base) | pretrained foundation LM | 44.05 avg (0-shot, our harness) |
| `HobbyLM-Chat.gguf` | [Chat](https://huggingface.co/rootxhacker/HobbyLM-Chat) | instruction / chat | 42.5 avg (alignment-tax dip from base) |
| `HobbyLM-Computer-Use.gguf` | [Computer-Use](https://huggingface.co/rootxhacker/HobbyLM-Computer-Use) | GUI agent + tool calling | 95% name-F1, 0% param-hallucination |
| `HobbyLM-Omni.gguf` | [Omni](https://huggingface.co/rootxhacker/HobbyLM-Omni) | multimodal core (text+image+audio) | VQAv2 47.0 / GQA 39.2 |
| `HobbyLM-Diffusion.gguf` | [Diffusion](https://huggingface.co/rootxhacker/HobbyLM-Diffusion) | masked-diffusion LM | 117 tok/s on H100 (~2.7× AR) |

Full benchmark tables, methodology, and limitations are on each model's own card (linked above).

## Running them

```bash
# from https://github.com/harishsg993010/HobbyLM
hobby-rs --model HobbyLM-Chat.gguf --prompt "The capital of France is" --n 48
```

## ⚠️ These use a custom `hobbylm` architecture

Every GGUF sets `general.architecture = hobbylm` (all metadata keys are `hobbylm.*`). **Stock llama.cpp will
not load them** — they need the from-scratch [`hobby-rs`](https://github.com/harishsg993010/HobbyLM) engine,
or a llama.cpp patched to register the `hobbylm` arch (GQA + per-head QK-norm + sigmoid-gated MoE + aux-free
routing bias + 1 shared expert + a leading dense layer). `HobbyLM-Diffusion` additionally carries `diffusion.*`
metadata and needs the diffusion-aware (bidirectional, iterative-denoise) decoder.

## License
Apache-2.0.