Tot-360M Family (GGUF) ? 345M Local AI Model Trained From Scratch
Official Project Hub & Live Raspberry Pi 4 Demo:
- Website: OB Studio (obstudio.org)
- Live Browser Demo (Running on a 4 GB Raspberry Pi 4): https://obstudio.org/tot-360m/demo/
- Full Technical Specification & Loss Curve: https://obstudio.org/tot-360m/
- Tot Desktop AI Companion for Windows: https://obstudio.org/tools/tot/
- Drift Privacy-First AI Browser: https://obstudio.org/tools/drift/
Overview
Tot-360M is a community-funded 345,359,296-parameter (314,639,296 non-embedding) decoder-only Transformer language model trained 100% from random initialization (from scratch) by Oliver at OB Studio across 8-core Kaggle Cloud TPU v5e-8 instances and Dual NVIDIA Tesla T4 GPUs.
Unlike fine-tuned Qwen or Llama derivatives, Tot-360M uses its own 32,000-token byte-level BPE tokenizer and was pre-trained and post-trained step-by-step in public. It exports directly to a standard Qwen3-compatible GGUF layout so that stock llama.cpp, LM Studio, PocketPal AI (iOS & Android), and ChatterUI can run it 100% offline with zero custom C++ patches.
At 241 MiB (Q4_K_M) and 351 MiB (Q8_0), Tot-360M runs smoothly on:
- Android & iPhone (under 420 MB RAM in PocketPal AI or ChatterUI)
- Raspberry Pi 4 & Pi 5 (4 GB / 8 GB RAM)
- Everyday Windows, macOS & Linux laptops without a discrete GPU (35+ tok/s on CPU)
Architecture & Hyperparameters
| Parameter | Value |
|---|---|
| Architecture | Decoder-only Transformer (Qwen3-compatible GGUF tensor layout) |
| Total Parameters | 345,359,296 (tied input/output embeddings) |
| Non-Embedding Parameters | 314,639,296 |
Layers (n_layers) |
32 |
Hidden Size (d_model) |
960 |
Intermediate (d_ff) |
2,560 (SwiGLU MLP) |
| Attention Heads | 15 Query / 5 KV (3:1 GQA, head_dim = 64, QK-Norm) |
| Positional Encoding | RoPE (theta = 10,000.0) |
| Normalization | RMSNorm (eps = 1e-6) |
| Vocabulary Size | 32,000 (byte-level BPE) |
| Context Window | 4,096 tokens (expanded from 1,024 in Run 4) |
| License | Apache-2.0 |
Training Runs & Checkpoints
All checkpoints and live telemetry are documented on the Tot-360M Technical Specification page and can be tested side-by-side on the Live Raspberry Pi 4 Demo:
| Checkpoint | Steps | Cumulative Tokens | Context | Loss | GGUF Downloads |
|---|---|---|---|---|---|
| Run 1 (Batch 1) | 439 ? 24,785 |
405,680,880 |
1,024 |
4.74 ? 2.73 |
Q4_K_M ? Q8_0 |
| Run 2 (Batch 2) | 24,785 ? 56,058 |
917,557,344 |
1,024 |
3.18 ? 2.81 |
Q4_K_M ? Q8_0 |
| Run 3 (Batch 3) | 56,058 ? 87,395 |
1,430,481,360 |
1,024 |
2.61 ? 2.03 |
Q4_K_M ? Q8_0 |
| Run 3 Part 2 (Irish-360M) | 87,395 ? 95,816 |
1,568,316,288 |
1,024 |
2.20 ? 2.00 |
run3_part2_1568m/ (in this repo) |
| Run 4 (4K Context Expansion) | 95,816 ? 101,591 |
1,662,910,788 |
4,096 |
1.39 ? 1.24 |
run4_instruct_4k/ (in this repo) |
| Run 5 (4K Base) | 101,591 ? 109,958 |
1,799,962,248 |
4,096 |
1.38 ? 1.14 |
Q4_K_M ? Q8_0 |
| SFT v2 (Post-Trained ? EN + Code) | 109,958 ? 113,958 |
1,865,482,248 |
4,096 |
2.64 ? 1.14 (masked) |
tot_360m_sft_v2_q4_k_m.gguf ? tot_360m_sft_v2_q8_0.gguf |
Quick Start (llama.cpp, LM Studio & PocketPal AI)
1. Command Line (llama.cpp)
./llama-cli -m tot_360m_sft_v2_q4_k_m.gguf -c 4096 \
-p "<|im_start|>user\nWrite a Python function to compute gravitational force.<|im_end|>\n<|im_start|>assistant\n" \
-n 256 --temp 0.7 --repeat-penalty 1.15
2. Android & iPhone (PocketPal AI / ChatterUI)
- Install PocketPal AI or ChatterUI on your phone.
- Search Hugging Face for
verycoololiver/Tot-360M-Familyor downloadtot_360m_sft_v2_q4_k_m.gguf(241 MiB). - Set context length to
4096and chat 100% offline in airplane mode. Full step-by-step phone & Raspberry Pi guide: How to Run a Free Offline AI Model on Your Phone, Laptop, or Raspberry Pi 4.
Engineering Build Logs & Post-Training Notes
- Tot-360M SFT v2 (1.87B) Is Live on a Raspberry Pi 4: Why We Switched to 100% English + Code
- How to Run a Free Offline AI Model on Your Phone, Laptop, or Raspberry Pi 4 (253 MB GGUF Guide)
- Tot-360M at 1.66B Tokens: 4K Context Expansion & Post-Training Pipeline
- Tot-360M Run 2 Complete (917M Tokens): Loss Curve & First GGUF Checkpoints
- Tot-360M's First Real Training Run: 405 Million Tokens on a Kaggle TPU
- More Browser Simulations, Games & Developer Tools at OB Studio
- Downloads last month
- 702
4-bit
8-bit
16-bit