Alieno/ailo-152m-v3

Applications

Claude Code ollama launch claude --model Alieno/ailo-152m-v3

Codex App ollama launch codex-app --model Alieno/ailo-152m-v3

OpenClaw ollama launch openclaw --model Alieno/ailo-152m-v3

Hermes Agent ollama launch hermes --model Alieno/ailo-152m-v3

Codex ollama launch codex --model Alieno/ailo-152m-v3

OpenCode ollama launch opencode --model Alieno/ailo-152m-v3

A 152M-parameter assistant, now with a better base model thanks to true logit-level knowledge distillation from a larger teacher — then instruction‑tuned. GGUF, runs on CPU and edge.

AILO-152M-v3 is the third iteration of AILO (Artificial Intelligence Language Operator). The novelty: instead of learning only from the teacher’s text, the base model was trained to match the teacher’s full token probability distribution (logit-KD, the classic Hinton KD) — possible because teacher and student share the same GPT-2 tokenizer. The improved base was then re-fine-tuned for chat, reasoning and tool use.

ollama run Alieno/ailo-152m-v3
>>> What is the capital of Italy?
The capital of Italy is Rome.

What’s new in v3


🧬 Method	true logit-KD (KL-divergence on token distributions) from GPT-Neo 1.3B (same GPT-2 vocab)
📉 Base LM	wikitext perplexity 126 → 84 (−33%) after distillation
💬 Assistant	chat perplexity −7.3% vs v2 (held-out, masked on responses) — then SFT (instruction + reasoning + tool)
🪶 Size	151.9M params · 97 MB (q4_k_m) – 291 MB (f16) · CPU & edge

Honest note: the teacher (GPT-Neo) is a raw 2020 LM, so logit-KD mainly improves base language modeling; the measured net effect on the final assistant is a real but modest −7.3% chat perplexity vs v2. Stronger teachers (e.g. Gemma) can’t be used for logit-KD here because their tokenizers differ from AILO’s.

Quick start (Ollama)

ollama run Alieno/ailo-152m-v3

Tags: :latest / :q8_0 (best, 156 MB) · :q4_k_m (smallest, 97 MB) · :f16 (291 MB).

Chat format

<|user|>
{question}
<|assistant|>
<think>{optional reasoning}</think>
{answer}<|end|>

Details

Property	Value
Parameters	151.9M
Architecture	Decoder-only Transformer (LayerNorm · RoPE · SwiGLU), 12L / 768 / 12H, ctx 512
Vocabulary	50,257 (GPT-2 BPE)
Pipeline	base → logit-KD from GPT-Neo 1.3B → SFT (Alpaca + GSM8K + SQuAD + tool-use)
Formats	GGUF (q4_k_m, q8_0, f16) — model only, no loader scripts

Limitations

152M params: limited world knowledge and multi-step reasoning. Best paired with retrieval/tools (the AILO system).
512-token context; short prompts. For exact math, use a calculator tool.
English. The logit-KD gain is bounded by the 152M capacity ceiling.

License & contact

Dual-license: CC BY-NC-SA 4.0 (research/education/personal) + commercial by separate agreement. Riccardo Sparacino — LinkedIn

@misc{ailo152m_v3_2026,
  title  = {AILO-152M-v3: A tiny LLM improved via true logit distillation},
  author = {Sparacino, Riccardo}, year = {2026},
  note   = {Dual-licensed CC BY-NC-SA 4.0 / commercial}
}

Acknowledgments

Teacher for logit-KD: GPT-Neo 1.3B (EleutherAI). Built with Ollama and llama.cpp.

AILO-152M-v3 Tiny LLM improved via true logit distillation ⚡