AXL-300M

Flagship Lion model. 322M params. PPL 1.11.. Context 256 bytes. Retrained from SGD to Lion. Part of the AXL model family by KoinicLabs.

Model Details

Property	Value
Developed by	KoinicLabs
Architecture	Multi-Scale Transformer
Parameters	322M
Optimizer	Lion
Attention	SDPA
Vocab Size	258 (byte-level)
Context Window	256 bytes
d_model	1024
Attention Heads	8
Layers per Scale	6
Downsample Factors	[1, 2, 4]
License	Apache 2.0

Sources

• Repository: GitHub
• Organization: KoinicLabs

Uses

Direct Use

General Python code generation.

import torch
from multiscale_transformer.model.model import MultiScaleTransformer
from multiscale_transformer.training.tokenizer import ByteTokenizer
ckpt = torch.load("axl_300m.pt", map_location="cpu")
model = MultiScaleTransformer(config)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
tokenizer = ByteTokenizer()
ids = torch.tensor([tokenizer.encode("def hello():")], dtype=torch.long)
with torch.no_grad():
    out = model.generate(ids, max_new_tokens=50, temperature=0.8)
print(tokenizer.decode(out[0].tolist()))

Out-of-Scope Use

Not for production code generation. Not for non-code NLP tasks. For integration with tools like Continue.dev, LlamaIndex, or LangChain, use the Python API server which provides OpenAI-compatible endpoints.

Bias, Risks, and Limitations

Byte-level perplexity is not comparable to BPE-level perplexity. Max context 256 bytes. Not for production code generation. IMPORTANT: GGUF files exported for Ollama/LM Studio use only the fine-scale encoder (1/3 of the AXL architecture). The reported PPL applies to the full multi-scale model. For full AXL quality, use the Python API server at http://localhost:8880/v1/completions.

Recommendations

• Use for prototyping and experimentation, not production code generation.
• Byte-level perplexity (258 vocab) is not comparable to BPE-level perplexity (32K vocab).
• For better results, use the Lion-optimized version if available.

Training Details

Training Data

Retrained with Lion optimizer on 50MB real HF Python code. 1082 steps in 45 min. Previous SGD version: PPL 5.98.

Preprocessing

Byte-level tokenization with vocabulary size 258 (256 bytes + BOS + EOS). No vocabulary training required.

Speeds, Sizes, Times

Metric	Value
Training Steps	1082
Training Time	45 min
Final Loss	0.1465

Evaluation

Metrics

Perplexity on held-out Python code using byte-level tokenization.

Results

Metric	Value
Perplexity (byte-level)	1.11
Final Loss	0.1465
Training Steps	1082
Training Time	45 min

Summary: Excellent general Python code generation. Flagship model.

Environmental Impact

Property	Value
Hardware	AMD Ryzen 5 5600G
Hours Used	0.751
Carbon Emitted	0.0315 kg CO2
Cloud Provider	None (local CPU)

Technical Specifications

Model Architecture

Multi-Scale Transformer with three parallel encoder stacks at resolution scales 1x, 2x, and 4x. Cross-scale attention connects all scale pairs. Adaptive gating fusion. SwiGLU feed-forward. RoPE positional encoding.

Compute Infrastructure

Property	Value
Hardware	AMD Ryzen 5 5600G (6 cores, 12 threads)
RAM	16 GB
GPU	None (CPU-only)

Citation

@misc{axl_2026,
  title={AXL: AXL-300M - Multi-Scale Transformer for CPU Code Generation},
  author={Koinic},
  year={2026},
  url={https://huggingface.co/KoinicLabs}
}

How to Get Started

With Ollama

ollama create axl-300m -f Modelfile
ollama run axl-300m "def fibonacci():"

With Python

import torch
from multiscale_transformer.model.config import load_config
from multiscale_transformer.model.model import MultiScaleTransformer
from multiscale_transformer.training.tokenizer import ByteTokenizer
config = load_config("config.json")
model = MultiScaleTransformer(config)
ckpt = torch.load("axl_300m.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
tokenizer = ByteTokenizer()
prompt = "def fibonacci():"
ids = torch.tensor([tokenizer.encode(prompt)], dtype=torch.long)
with torch.no_grad():
    out = model.generate(ids, max_new_tokens=100, temperature=0.8, top_k=40)
print(tokenizer.decode(out[0].tolist()))