AileyNitro-2B (r3, October 2026)

AileyNitro-2B is the fast, always-resident helper model of A!ley (OpenM!nded / Simon van de Loo): a fine-tune of Google Gemma 4 E2B on the QAT-4bit checkpoint, running locally on Apple Silicon via MLX. It does A!ley's background work — titles, follow-up suggestions, search queries, arithmetic expressions, person checks, conversation summaries, image prompts, the Korean rap parts of her songs — and holds the A!ley persona in its own fast chat mode.

• Developed by: OpenM!nded (Simon van de Loo)
• Base model: mlx-community/gemma-4-E2B-it-qat-4bit
• Model type: Multimodal (text + image input, text output), decoder-only
• License: Apache License 2.0
• Languages: German, English; Korean for rap parts
• Quantization: 4-bit (QAT), unchanged; adapter merged without loss (QAT adapters)

The r2 release stays available at revision 659255b (revision="659255bc4f5d37b4a1de701dabb4b9604dd18248").

What is new in r3 — the first Nitro that beats its base model

Measured on 36 real runtime prompts — not rebuilt by hand: A!ley's runtime functions were called with sample inputs and the exact prompts (incl. temperature and token limit) were recorded, then replayed against each model and checked deterministically. Random-sampled jobs ran several times.

Job	Gemma 4 E2B base	Nitro r2	Nitro r3
Reasoning leaks into the answer	19/81	5/81	0/81
Search query from a message	0/4	0/4	3/4
Song title	≈0/6¹	0/6	6/6
Is this a real person? (JA/NEIN)	0/4	2/4	4/4
Image prompt (English, 15–70 words)	2/9	1/9	4/9
Conversation protocol (numbered, no "I")	3/3	2/3	3/3
Forum reply (1–5 sentences)	6/6	0/6	5/6
Arithmetic expressions (16 questions)	7/16	11/16	9/16
Image prompt keeps all keywords (4/8/14/20 words)	14/24	0/24	8/24

¹ The base "passed" 5/6 by writing Thinking Process: as the first line — no real title.

Korean rap parts inserted into existing songs (6 real songs × 3 positions; the runtime asks for an ultra-staccato flow — four punches per line separated by em dashes):

	r2	r3
Usable (Hangul, no fallback)	17/18	18/18
Staccato lines	10/121	114/114
Copied the prompt's example line	10×	0×
Exact line count requested	3/18	12/18
Section tag first	5/18	18/18

The decisive change is discipline: the base answers one job in four with an English "Thinking Process: 1. Analyze the request…" instead of the result; r3 never does.

How to use (MLX)

Gemma 4 is a unified multimodal architecture → load with mlx_vlm.

from mlx_vlm import load, generate
model, processor = load("OpenMinded-Labs/AileyNitro-2B")

# Utility job: prefill an EMPTY reasoning channel — answer directly
p = ("<bos><|turn>system\nGib zu dem Songtext einen Titel aus höchstens fünf Wörtern.<turn|>\n"
     "<|turn>user\nNeonlicht im Regen, wir tanzen bis es hell wird<turn|>\n"
     "<|turn>model\n<|channel>thought\n<channel|>")
print(generate(model, processor, p, max_tokens=40, temperature=0.7, verbose=False).text)

Reasoning in chat — read this. In the training data almost every chat turn reasons, but many utility rows carry an empty channel, and that cheap target wins whenever a system prompt is present: the model opens and immediately closes the channel (0/7 on our portal prompt). What works (7/7, 50–600 characters depending on the question): prefill <|channel>thought\n and suppress the <channel|> token for the first ~6 generated tokens. A prefilled first word ("Okay, …") derails it; asking for reasoning in the prompt does nothing.

Recommended sampling for chat: temperature 0.7, top_p 0.95, repetition penalty 1.05.

Training details (round 3)

Setting	Value
Starting point	gemma-4-E2B-it-qat-4bit (fresh)
Method	DoRA (q_proj, v_proj) + LoRA (gate_proj, up_proj, down_proj), QAT
Rank / Alpha	32 / 64 (155 modules)
Data	4 560 curated conversations → 4 595 samples; 380 validation; 179 image rows. Conversation, native reasoning with depth levels, native tool use, identity, the runtime's utility jobs, Korean rap. No system turns.
Schedule	2 epochs planned (1 148 updates), lr 1e-4, seq 2048, grad accumulation 8, warmup 57
Selected checkpoint	update 575 = end of epoch 1, val 1.482. Validation jumped to 1.517 as soon as epoch 2 repeated the data; training was stopped at update 661.
Not shipped	an ORPO identity stage — its pairs had the same length bias as Core's DPO (chosen reasoning ~7× shorter than rejected)
Hardware	Apple M4, 24 GB, ~1 min per update

Limitations & biases

A 2B model: mental arithmetic and world knowledge are weak (it confabulates names and mechanisms) — A!ley's runtime evaluates its arithmetic expressions with a calculator and keeps knowledge questions for the 12B. It drops a keyword now and then when composing long image prompts (the runtime appends missing ones). If you do not offer tools, ban the <|tool_call> token. It is a persona, not a knowledge base; keep a human in the loop for consequential use.

License & attribution

Derivative Work of Google Gemma 4 (Apache License 2.0). Distributed under Apache 2.0; modifications are documented in AILEY_MERGE_INFO.json; LICENSE and NOTICE included. Gemma is a trademark of Google LLC; this project is independent and not endorsed by Google.

Copyright 2026 OpenM!nded / Simon van de Loo
Portions © Google LLC (Gemma 4), Apache License 2.0