Instructions to use Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories

SGLang

How to use Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories with Docker Model Runner:
```
docker model run hf.co/Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories
```

Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories

THOX Microprocessor is an attention-dominant tiny decoder model designed for local-first routing, structured extraction, device control, and compact on-device language workloads.

Architecture

Property	Value
Parameters	8,093,520
Non-embedding parameters	5,472,080
d_model	320
Layers	21
Attention heads / KV heads	8 / 2
Context	1024
QK RMS normalization	True
Sandwich normalization	True
Blockwise Walsh-Hadamard mixer	True
Architecture version	thox-microprocessor-v0.2
Trained tokens	98.3M (TinyStories)

Weight payload at INT4 (group size 64, including per-group scales and the fp32 1-D tensors): 4.45 MB, i.e. 4.40 effective bits per parameter. This is the raw weight payload, not peak runtime RAM — device qualification must additionally measure code, allocator overhead, activations, and KV cache.

Status

Architecture-validation checkpoint — not a product model. Held-out loss 1.6956 / perplexity 5.4499 on the TinyStories validation split (2,095,104 target tokens). THOXQ4 INT4 costs +0.0366 held-out loss at 4.45 MB; INT4 with an INT8 embedding costs +0.0162 at 5.76 MB. Trained ONLY on TinyStories (roneneldan/TinyStories, CDLA-Sharing-1.0 upstream) — simple synthetic children's stories. It has NO routing, tool-call, device-control or THOX domain capability, and it is NOT the 8M Control Core.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxMicroprocessor-8M-Base-TinyStories", trust_remote_code=True)

The remote-code reference implementation recomputes the prefix on each decoding step rather than maintaining a KV cache; it is optimised for portability across transformers versions. Use the thox_microprocessor package for cached decoding.

Downloads last month: 11

Safetensors

Model size

8.09M params

Tensor type

F32