ThoxLLM-327M-v2
Your AI. Your Data. Your Rules.
Production from-scratch 327M decoder with standard 2-projection SwiGLU (GGUF-clean).
What this is
- Trained from scratch (THOX-owned) — Apache-2.0.
- Standard SwiGLU (matched-FLOPs, ffn 3413) makes GGUF export clean via K-quants.
- config.json has no
architectures/model_typekey, so it is not vanilla-AutoModelloadable — load via the THOXthoxllmpackage or use the GGUF repo. - Dataset provenance note: parent card says THOX-internal corpus; the GGUF card says Pile-uncopyrighted — reconcile before publishing a data statement.
Architecture (from config)
| Field | Value |
|---|---|
| Architecture | custom THOX decoder (standard SwiGLU; no HF architectures key) |
| Layers | 20 |
| Hidden size | 1024 |
| Attention heads | 16 |
| KV heads | 8 (GQA) |
| FFN / intermediate | 3413 |
| Vocab | 50,304 |
| Max context | 2048 |
| Tied embeddings | yes |
| dtype | bfloat16 |
Intended use
On-device / edge text generation within the THOX stack. Not a safety-aligned public assistant unless deployed behind THOX guardrails.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Thox-ai/ThoxLLM-327M-v2")
model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxLLM-327M-v2")
Note: this checkpoint uses the custom THOX thoxllm decoder; config.json has no architectures key, so vanilla AutoModel loading may not work — use the THOX thoxllm package or the GGUF build.
Links
- GGUF:
Thox-ai/ThoxLLM-327M-v2-GGUF - Ollama:
ollama.com/thox-ai/<slug>— verify with the Ollama lane (task 80017303) - Docs: https://docs.thox.ai
THOX.ai LLC — Your AI. Your Data. Your Rules. · On-device and private by design.
Temporary demo Space: tommytracx/ThoxLLM-327M-v2-Space (org Space paused on quota).