Instructions to use CastIronMind/Lissa-1.5B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use CastIronMind/Lissa-1.5B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="CastIronMind/Lissa-1.5B-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("CastIronMind/Lissa-1.5B-Instruct", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use CastIronMind/Lissa-1.5B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "CastIronMind/Lissa-1.5B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CastIronMind/Lissa-1.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/CastIronMind/Lissa-1.5B-Instruct

SGLang

How to use CastIronMind/Lissa-1.5B-Instruct 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 "CastIronMind/Lissa-1.5B-Instruct" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CastIronMind/Lissa-1.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "CastIronMind/Lissa-1.5B-Instruct" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CastIronMind/Lissa-1.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use CastIronMind/Lissa-1.5B-Instruct with Docker Model Runner:
```
docker model run hf.co/CastIronMind/Lissa-1.5B-Instruct
```

Lissa 1.5B: a 1.45B model trained from scratch for $555, with the full cost accounting

by CastIronMind - opened about 21 hours ago

Discussion

CastIronMind

Owner about 21 hours ago

Opening this thread so the numbers live somewhere other than the model card.

What this is: a 1.45B parameter language model trained from scratch (no distilled weights, no borrowed backbone) on a single rented RTX 5090 for about $555. 30B pretraining tokens, roughly 21 per parameter, which is Chinchilla-optimal for that budget. Hybrid of Kimi Delta Attention and grouped-query attention: 18 linear and 6 full attention layers out of 24, plus attention on residuals. Then supervised on 59,182 examples.

What I want this release to be useful for: an honest answer to "what does $555 of pretraining buy in 2026". So the card keeps the results that hurt. HellaSwag 29.95 (Pythia-1.4B: 32.35), ARC-Easy 37.63 (45.58), PIQA 60.39 (66.05), MMLU 23.03, IFEval prompt-strict 17.19. BPC on held-out text 0.8911 English, against 0.4009 for Qwen2.5-1.5B, which is what 30B tokens versus 18T tokens looks like.

Where it holds up: multi-turn roleplay and character work. 91 of 100 behavioural samples terminate cleanly on , 30 of 30 roleplay prompts stay in character with no self-name leakage, 17.8 tokens/s at batch 1, 4.57 GB peak VRAM at a 16,384-token prompt.

Open questions I would genuinely like answered by anyone who has run this experiment at a larger scale: at what token count does the gap to web-text models at 1.5B actually close, and is the crossover dominated by data quality or by token count?

The GGUF route does not exist yet, the architecture needs fla 0.5.2 and Triton. If there is interest, a Llama-architecture distillation of this checkpoint is the obvious next step, and the 103M and 32M models in the same family already run in llama.cpp.

CastIronMind

Owner about 17 hours ago

What am I doing with my life... Maybe I should try some shitty fine-tuning of a model—something along the lines of "UNCENSORED CLAUDE CHATGPT TURBO-NANO!!!!1111"

CastIronMind

Owner about 17 hours ago

One like and i train JESUS CHRIST GPT only on bible and some another books

CastIronMind

Owner about 16 hours ago

Or furry femboy vae will be better?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment