Skip to content
Other

Hedwig 97.6M: A Compact Model for Stories and Grammar Editing

3 min readupdated

Hedwig 97.6M is a miniature 97.6M-parameter ChatGPT-style language model (LLM) built and trained from scratch using pure PyTorch, without high-level frameworks like HF Trainer or Lightning. It can write short English fairy tales and correct grammar and tone in casual text. Based on a modern Transformer architecture (12 layers, RoPE, RMSNorm, SwiGLU, SDPA) with a custom 8,192-token BPE tokenizer, it goes through pre-training, SFT (instruction fine-tuning), and Anchored DPO (preference alignment). Thanks to its small footprint and quantized versions (8-bit and 4-bit weight compression), the model runs completely offline even on laptops—achieving 54 tokens per second on an Apple M4 chip.

Key features

  • Built from scratch in PyTorch: The entire pipeline—from the tokenizer and pre-training on 159M tokens (TinyStories, FineWeb-Edu, and Project Gutenberg classics) to DPO and quantization—is implemented without third-party wrappers.
  • Extreme compactness: In the base bfloat16 format, the checkpoint weighs 195.6 MB, while the int8 and int4 quantized versions take up just 98.2 MB and 55.2 MB, respectively.
  • Two primary modes: The model is fine-tuned on 30,930 instruction pairs from the TinyStories-Instruct and CoEdIT datasets for story generation and basic editing tasks.
  • Fast local inference: Optimized for offline execution on consumer hardware, including Apple Silicon Macs.

What you can do

  • Local story generator: Embed the checkpoint into an offline application or script to generate simple English stories without internet access.
  • Basic short-text correction: Use it for quick grammar and style edits in short, everyday messages.
  • LLM architecture study: Explore the open-source training code in pure PyTorch and visualize layer activity using the included Brain X-Ray UI.

Example

Prompt: Write a story about a boy named Harry Potter.
Model response (54 tok/s offline on Apple M4): Harry Potter was a humble boy who liked to play. One day, he decided to go out and play. He walked down the street, laughing and having so much fun. Suddenly, Harry saw something shiny in a store window...

Read moreDetails
Metrics: The final brain_dpo.pt checkpoint showed a Val/Test Loss of 1.235 and a 95.0% saved win-rate during the DPO stage (peaking at 96.7%); for the int4 version, the loss was 1.4118.
Training: Took about 66 minutes on a single NVIDIA A40 GPU (48 GB) and cost roughly $0.65.
Limitations: With 97.6M parameters, this is an educational model for stories rather than a general-purpose assistant; in long work emails, it may cut off text or miss tense agreement.
Additional info: The release also introduces a larger hedwig-300m variant, and the source code is available in the author's GitHub repository.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.2 GB · ready files0.7 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ansarbuilds/hedwig-97m
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ansarbuilds/hedwig-97m --local-dir ~/models/hedwig-97m
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to hedwig-97m
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of hedwig-97m (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

ansarbuilds/hedwig-97m
Author
ansarbuilds
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments