# Running Cosmos — Atomic, Ollama, llama.cpp, or plain Python

**Read this first, because it will save you an hour.** Cosmos is not a fine-tune of an
existing model. Its attention runs a mechanism (`LLM_ARCH_COSMOS`) that no released
inference runtime knows about, and `llama.cpp` dispatches on `general.architecture`
against an enum compiled into the binary. So:

| you want | works today | why |
|---|---|---|
| **Talk to her right now, no build** | ✅ `serving/cosmos_serve.py` | PyTorch, speaks the Ollama HTTP API |
| **Her architecture natively in C++** | ✅ the `cosmos-arch` llama.cpp fork | has the enum entry |
| **Stock Atomic / stock Ollama / LM Studio** | ❌ not yet | their bundled llama.cpp has no `LLM_ARCH_COSMOS` |

No Modelfile, flag, or conversion changes the third row. It needs a runtime built with
the architecture registered. That is a pull request to those projects, not a setting.

---

## 1. Fastest path — talk to her in about a minute

Needs Python 3.9+ and `torch`. Nothing else.

```bash
pip install torch
python serving/cosmos_serve.py 11501
```

That serves **PHOS**, **CST** and **Spark** behind the Ollama HTTP API on port 11501, so
any Ollama-compatible client works by pointing at it instead of 11434:

```bash
curl http://127.0.0.1:11501/api/tags
curl http://127.0.0.1:11501/api/chat -d '{"model":"cosmos-phos","messages":[{"role":"user","content":"hello"}],"stream":false}'
```

In Atomic, LM Studio, or anything else that accepts a custom Ollama endpoint, set the
base URL to `http://127.0.0.1:11501`. The models appear by name.

Or just use the launcher, which starts the server and waits for the port:

```
START_COSMOS_KIT.bat     ->  1. Chat with Cosmos
```

---

## 2. Native C++ — her real architecture in llama.cpp

`cosmos-cst.gguf` in this repository is her 54D Mixture-of-States Hebbian attention as a
real GGUF: 61 tensors, `general.architecture: cosmos`, `attention.key_length = 102`
(head_dim 48 + the 54-dimensional state riding in the key rows).

Stock llama.cpp will refuse it — correctly, because it does not have the architecture.
Build the fork:

```bash
git clone -b cosmos-arch https://github.com/NavisWORLD/llama.cpp
cd llama.cpp
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_SERVER=OFF -DLLAMA_CURL=OFF
cmake --build build --config Release --target llama-completion
./build/bin/Release/llama-completion -m cosmos-cst.gguf -p "the woods" -n 48 -c 128
```

Measured on a 6-core CPU: **2,411 tok/s prompt eval, 1,078 tok/s generation.**

> This is a **1.9M-parameter character-level model.** It produces proto-English with real
> structure (`a dimly lit ... bedroom room`) and is not fluent. That is the expected
> result at this size and is not a defect. If you want fluency, this is the wrong repo;
> if you want an architecture nobody else has running, this is it.

---

## 3. Plain PyTorch

```python
import torch
ck = torch.load("weights/phos.pt", map_location="cpu", weights_only=False)
print(ck["arch"], f'{ck["total_steps"]:,} steps', ck["best_val_loss"])
# PHOS-dyn12-phi-QuantumBorn 2,800 steps 0.6984356045722961
```

`weights/phos.pt`, `weights/cosmos_born.pt` and `weights/spark_cst.pt` are all
architecturally original and quantum-born. The model classes live in `architecture/`.

---

## What changed on 2026-08-01, and why your old bookmark broke

Earlier versions of this file told you to download `cosmos-namebind-weights.gguf`. **That
file was removed.** It was a Qwen2.5-1.5B fine-tune, not this architecture — its GGUF
metadata carried `general.architecture: qwen2`, which this hub rendered as an
architecture badge above the model card, so the page announced someone else's base model
as this work's architecture. Removing it also made the card's own statement true: this
repository distributes only weights that are architecturally original.

If you specifically want that conversational teacher model, it is a Qwen2.5-1.5B
derivative under **Apache 2.0** © Alibaba Cloud, and you should get Qwen from Alibaba
rather than from here. See §1b of the [model card](../README.md) for the full provenance.