        v5.0 — open, local-first
      

PorchLM

A small ~50M-parameter language model you can train from scratch and run entirely on your own machine — LM Studio, Ollama, llama.cpp, or right here in the browser. Nothing you type ever has to leave your device.

Every number below is measured, not estimated. The exported model is checked against the official GGUF reader, and its compute graph has been independently re-implemented in NumPy — logits match PyTorch to about 1e-6.

Try it in chat Download What is this?

Specifications

Parameters

~50M

trained weights

Context

4,096

tokens, RoPE

Size

~137 MB

extremely small

License

MIT

runs 100% locally

Figures for v5.0, the most recent release. Earlier versions differ — each changelog entry lists its own.

Architecture

Architecture Llama (RoPE, RMSNorm, SwiGLU, GQA, tied embeddings)

Layers 10

Embedding dim 640

Attention heads 8 query / 4 key-value (GQA)

Head dim 64

Feed-forward dim 1728 (SwiGLU)

Context length 4,096 (RoPE, base 10000)

Normalization RMSNorm, eps 1e-5

Vocabulary 16,384 (byte-level BPE)

Output head Tied to token embeddings

Chat format ChatML (<|im_start|> / <|im_end|>)

Performance

Browser inference ~150–500 tok/s (CPU only)

Training time ~5 minutes total (Colab T4)

F16 file size ~16 MB

Q4_K_M file size ~6 MB

Cold load in browser ~1 second for 16 MB, local file

Measured on a Colab T4 and a typical laptop CPU — your numbers will vary.

Recent updates

v5.0 Coming soon 20B tokens on a 50M parameter model. Less Wikipedia, more stories, a smarter structure overall. v5.0 Coming soon 20B tokens on a 50M parameter model. Less Wikipedia, more stories, a smarter structure overall. v4.0 Shipped An experimental, smaller, and unfortunately dumber past model.

See the full changelog