GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
pinkelephantlimited
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
pinkelephantlimited
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Parameter | Value |
|---|---|
| Architecture | LLaMA (decoder-only) |
| Parameters | 66,268,416 |
| Hidden size | 768 |
| Intermediate size | 2,304 |
| Num layers | 7 |
| Num attention heads | 12 |
| Num KV heads | 12 |
| Max sequence length | 1,024 |
| Vocabulary size | 16,384 |
| Position encoding | RoPE (θ=10000.0) |
| Activation | SwiGLU |
| Normalization | Pre-RMSNorm |
| Weight tying | Yes (embedding ↔ lm_head) |
| Precision | bf16 |
| # | Dataset | Domain | Max examples |
|---|---|---|---|
| 1 | FineWeb-Edu (sample-10BT) | Educational web text | 60,000 |
| 2 | FineWeb (sample-10BT) | General web text | 60,000 |
| 3 | OpenWebMath | Mathematical text |
| Parameter | Value |
|---|---|
| Hardware | 96GB GPU |
| Optimizer | AdamW 8-bit (bitsandbytes) |
| Learning rate | 2e-4 |
| Batch size | 128 |
| Warmup steps | 500 |
| Weight decay | 0.01 |
| Save interval | 500 steps |
| Checkpoint upload | Auto (instant to HF) |
python
from transformers import pipelinepipe = pipeline("text-generation", model="pinkelephantlimited/pink-elephant-66m")output = pipe("Photosynthesis is the process by which", max_new_tokens=80)[0]["generated_text"]print(output)
This model was trained entirely from scratch on permissively licensed data. No fine-tuning, no transfer learning — 100% original weights.
MIT
python
from transformers import pipelinepipe = pipeline("text-generation", model="pinkelephantlimited/pink-elephant-66m")pipe("The water cycle consists of three main stages:", max_new_tokens=80)
python
pipe("In conclusion, the key factors that led to", max_new_tokens=60)
python
pipe("def fibonacci(n):", max_new_tokens=60)
Why 66M? 768-dimensional hidden size with 7 layers — deeper and wider than 55M, giving noticeably better quality while remaining CPU-friendly.
| 40,000 |
| 4 | SmolLM Corpus (cosmopedia-v2) | Educational books | 40,000 |
| 5 | CodeParrot | Source code (multi-language) | 30,000 |
| 6 | Nemotron-Legal (Case Law) | Legal case law | 30,000 |
| 7 | Investopedia | Financial terms | 15,000 |
| Total | ~275,000 |