minnow-pico-50M-base
A 48.5M parameter DeepSeek-V4 architecture model trained from scratch on a multilingual blend of C4 data (108 languages). This is a research sandbox model for studying the DeepSeek-V4 architecture at small scale — it is not a production model and has no practical use.
Architecture
Based on the DeepSeek-V4 architecture,
consumed via HuggingFace DeepseekV4ForCausalLM (transformers ≥ 5.14):
- 256-dim hidden, 8 layers, 4 heads × 64 head_dim
- 16 experts, top-2 routing (Sqrt(Softplus) affinity)
- Hash-routed MoE on first 3 layers, learned routing on last 2
- Hybrid attention (CSA + HCA + sliding)
- Manifold-Constrained Hyper-Connections (mHC)
- Grouped low-rank attention output projection
- 48,538,724 parameters total
Training
- Data: 1B tokens from a multilingual C4 blend (108 languages)
- Optimizer: AdamW (betas 0.9, 0.95)
- Learning rate: 6e-4, cosine schedule, 763 warmup steps
- Batch size: 131,072 tokens/step (2× RX 7900 XTX, NCCL DDP)
- Steps: 7,629
- Final loss: 3.49
- Hardware: 2× AMD RX 7900 XTX (gfx1100, RDNA3)
This model was trained from scratch (random initialization) — it does not inherit weights from any pretrained model.
Tokenizer
Uses the DeepSeek-V4-Flash-0731 tokenizer (vocab 129,280).
Disclaimer
This model exists purely for architectural experimentation. At 48.5M parameters, it cannot perform any useful tasks. Do not use it for anything.