AXL — Architecture eXperimental Lab
27 CPU-Optimized Code Generation Models by Koinic
All models are trained from scratch on consumer hardware (AMD Ryzen 5 5600G, 16GB RAM). No GPU required.
Models
Lion-Optimized (Recommended)
| Model | Params | PPL | GGUF (F16) | GGUF (Q4_K_M) |
|---|---|---|---|---|
| AXL-Code-1B-Lion | 318M | 1.90 | 606 MB | 188 MB |
| AXL-Reasoning-Lion | 70M | 1.03 | 134 MB | 44 MB |
| AXL-Refactor-Lion | 19.1M | 1.02 | 37 MB | 12 MB |
| AXL-TestGen-Lion | 15.2M | 1.02 | 30 MB | 18 MB |
| AXL-Chat-Lion | 9.9M | 1.03 | 19 MB | 7 MB |
| AXL-Micro-Lion | 12.8M | 1.04 | 25 MB | 15 MB |
| AXL-Secure-Lion | 11.7M | 1.03 | 23 MB | 8 MB |
| AXL-Docs-Lion | 9.9M | 1.01 | 19 MB | 7 MB |
| AXL-Comment-Lion | 7.2M | 1.02 | 14 MB | 5 MB |
SGD Models
| Model | Params | PPL | Focus |
|---|---|---|---|
| AXL-Micro-600K | 600K | 63.08 | Demo |
| AXL-Micro-8M | 12.8M | 3.13 | Code gen |
| AXL-Coder-15M | 26.0M | 5.97 | Agentic |
| AXL-Debugger-8M | 14.1M | 6.60 | Bug fixing |
| AXL-Fixer-12M | 20.9M | 5.90 | Debug |
| AXL-Reasoning-70M | 70M | 1.93 | CoT |
| AXL-300M | 322M | 5.98 | Flagship |
| AXL-Chat-10M | 9.9M | 1.02 | Dialogue |
| AXL-TestGen-15M | 15.2M | 1.01 | Test gen |
| AXL-Refactor-20M | 19.1M | 1.01 | Refactoring |
| AXL-Docs-8M | 9.9M | 1.03 | Docstrings |
| AXL-Comment-5M | 7.2M | 1.01 | Comments |
| AXL-Secure-10M | 11.7M | 1.01 | Security |
Specialized Models
| Model | Params | PPL | Focus |
|---|---|---|---|
| AXL-Code-1B | 318M | 31.22 | Code gen (SGD) |
| AXL-Chat-Pro | 12.8M | 3.42 | Advanced chat |
| AXL-Translate | 15.2M | 1.86 | Code translation |
| AXL-Vision-0.8M | 1M | — | Vision encoder |
| AXL-Vision-v2 | 4.1M | — | UI vision |
Quick Start
Python API Server (Full Quality - Recommended)
pip install -e .
python AXL/API/serve_model.py --model AXL-Micro-Lion/ --port 8880
# OpenAI-compatible endpoint:
curl http://localhost:8880/v1/completions \
-H "Content-Type: application/json" \
-d '{"prompt": "def fibonacci(n):", "max_tokens": 100}'
# Works with Continue.dev, LlamaIndex, LangChain, Cursor
With Ollama (Degraded Quality)
Warning: GGUF files for Ollama use only the fine-scale encoder (1/3 of the AXL architecture). The reported PPL values apply to the full multi-scale model. Use the Python API above for full quality.
cd AXL-Micro-Lion
ollama create axl-micro-lion -f Modelfile
ollama run axl-micro-lion "def fibonacci(n):"
With Python (Direct Inference)
import torch
from multiscale_transformer.model.config import load_config, ModelConfig
from multiscale_transformer.model.model import MultiScaleTransformer
from multiscale_transformer.training.tokenizer import ByteTokenizer
config = load_config("AXL-Micro-Lion/config.json")
model = MultiScaleTransformer(config)
ckpt = torch.load("AXL-Micro-Lion/axl_micro_lion.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
tokenizer = ByteTokenizer()
ids = torch.tensor([tokenizer.encode("def hello():")], dtype=torch.long)
out = model.generate(ids, max_new_tokens=50, temperature=0.8, top_k=40)
print(tokenizer.decode(out[0].tolist()))
Architecture
AXL processes token sequences at three parallel resolution scales:
- Fine (1x): All tokens. Attention cost: O(N^2 d)
- Medium (2x): Grouped in pairs. Cost: O(N^2 d/4)
- Coarse (4x): Grouped in quadruplets. Cost: O(N^2 d/16)
Cross-scale attention connects all scale pairs. Adaptive gating fusion combines representations.
Lion optimizer: Sign-based momentum, 20x faster convergence than SGD, 50% less memory than AdamW.
Byte-level tokenizer: 258 vocab (256 bytes + BOS + EOS). No vocabulary training. Works with any programming language.
Training Cost
| Model | Time | Cost (USD) |
|---|---|---|
| AXL-Comment-Lion | 2 min | $0.0004 |
| AXL-Code-1B-Lion | 20 min | $0.004 |
| All 9 Lion models | 49 min | $0.010 |
Based on AMD Ryzen 5 5600G (100W system, $0.12/kWh).
Papers
Code
Full training code: GitHub