AileyCore-12B (r3, October 2026)

AileyCore-12B is a fine-tune of Google Gemma 4 (12B, instruction-tuned) on Google's QAT 4-bit checkpoint, built to run locally on Apple Silicon via MLX. It powers A!ley, the on-device persona created by OpenM!nded / Simon van de Loo, and is her default mode in the A!ley portal and app.

• Developed by: OpenM!nded (Simon van de Loo)
• Model type: Multimodal (text + image + audio input, text output), decoder-only
• Base model: mlx-community/gemma-4-12B-it-qat-4bit
• License: Apache License 2.0
• Languages: German, English (answers in the user's language)
• Quantization: 4-bit QAT (mixed: some MLP layers 8-bit, as in the base); adapter merged without loss

r3 is a plain merged checkpoint — no side path, no extra loader code. The r2 release (6-bit with a rank-8 side path) stays available at revision 37124f0 (revision="37124f02f70770d62522af9be15e3d440f737dc0").

---

What changed in r3

r1 and r2 were successive corrections on the same 6-bit weights. r3 starts fresh from Google's QAT checkpoint and trains everything A!ley needs in one go, all in Gemma 4's native formats:

• Native reasoning channel (<|channel>thought … <channel|>) with depth scaled to the task — a greeting gets a sentence, a puzzle gets a plan.
• Native tool use: Gemma 4 tool declarations, <|tool_call> calls and <|tool_response> results exactly as the runtime produces them (web search, RAG over her own songs, books and projects, image generation, recipes, calculator …), including when not to call a tool.
• Identity without a prompt: 81 % of the training rows have no system turn at all; the user's name sits in the user turn (Name: message).
• Narrative / prose writing, vision rows (441 images), multilingual answers.

QAT adapters — why the merge is lossless

The adapters were trained as QAT adapters: the forward pass sees fake_quant(W_dequant + ΔW) with a straight-through gradient, reading bits and group size per layer (so the base's mixed 4/8-bit layout is respected). The model only learns deltas that survive the rounding, and fuse() yields weights bit-identical to what was trained. (In r2, a classic merge into 6-bit had rounded away 87 % of the delta — hence the side path back then.)

Measured (isolated, before going live)

Benchmark run with the model loaded alone, same questions for every condition:

	no system prompt	one-line context header	runtime prompt (~700 chars, tool hints)	old long persona prompt (~7.8k chars)
Logic puzzles	5/5	4/5	3/5	1/5
Reasoning block never closed	1/8	1/8	2/8	4/8
Depth scaled to task	6/6	6/6	—	4/6
Tool: right call / correctly none	8/10 · 4/6	7/10 · 5/6	9/10 · 5/6	9/10 · 5/6
Answers in the user's language (es/en/fr/tr/ja)	—¹	4/5	5/5	4/5

¹ Without any prompt the model does not know a portal visitor is writing and answers in German.

Identity with no prompt at all: 11/12 answers as A!ley, 0/12 as another model.

The important finding for anyone using this model: long system prompts hurt it. Because it learned identity and behaviour without one, a long prompt makes the reasoning collapse (shorter, unclosed, wrong). Use no system prompt or a short one.

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/AileyCore-12B")
prompt = "<bos><|turn>user\nSimon: Wer bist du?<turn|>\n<|turn>model\n"
print(generate(model, processor, prompt, max_tokens=600, temperature=0.65,
               top_p=0.85, repetition_penalty=1.15, verbose=False).text)

Answers come in the native channel format:

<|channel>thought
…reasoning, as long as the task needs…
<channel|>…answer…<turn|>

Recommended sampling (measured, not the Gemma defaults): temperature 0.65, top_p 0.85, repetition penalty 1.15. For tool use pass the tool schemas to apply_chat_template(tools=…) and return results directly after <tool_call|> as <|tool_response>response:NAME{value:<|"|>…<|"|>}<tool_response|>.

Training details (round 3)

Setting	Value
Starting point	gemma-4-12B-it-qat-4bit (fresh — not on top of r1/r2)
Method	DoRA (q_proj, v_proj) + LoRA (gate_proj, up_proj, down_proj), QAT
Rank / Alpha	16 / 32 (232 modules, 55 M trainable parameters)
Data	2 429 curated conversations → 2 676 samples after chunking long documents; 268 validation conversations (296 samples)
Content	conversation, native reasoning with depth levels, native tool calls/responses, RAG/web context as tool results, identity, prose, 441 image rows
Epochs / updates	1 / 334 (grad accumulation 8, seq 2048, lr 2e-5, warmup 16, gradient checkpointing)
Validation loss	1.692 (update 25) → 1.318 (best, update 325)
Not shipped	a DPO stage: in the preference pairs the chosen reasoning was ~9× shorter than the rejected one (137 vs 1 228 chars), so it taught "think short", not "think right"
Hardware	Apple M4, 24 GB unified memory
Framework	MLX (mlx_vlm + mlx_lm.tuner), QAT layers from the A!ley training code

Intended use

• Local, privacy-respecting assistant on Apple Silicon Macs
• Conversational reasoning, writing, tool-using assistance in DE/EN (and the user's language)
• Multimodal understanding (image / audio input) inherited from Gemma 4

Out of scope

• Any use prohibited by applicable law
• Safety-critical, medical, legal, or financial decision-making without human oversight

Limitations & biases

Inherited from Gemma 4 plus the fine-tune: it can produce inaccurate, outdated or biased content; it is a persona, not a knowledge base — let it search when facts matter. With long system prompts reasoning degrades (see above). Keep a human in the loop for consequential use.

License & attribution

This model is a Derivative Work of Google Gemma 4, released under the Apache License 2.0 (see the official Gemma 4 license). AileyCore-12B is distributed under Apache 2.0.

In accordance with Apache 2.0 §4: the base weights were modified via DoRA/LoRA adaptation (QAT adapters, merged — see AILEY_MERGE_INFO.json); a copy of the license is included (LICENSE); attribution notices are in NOTICE.

Gemma is a trademark of Google LLC. This project is independent and not endorsed by or affiliated with Google.

Copyright 2026 OpenM!nded / Simon van de Loo
Portions © Google LLC (Gemma 4), Apache License 2.0
Licensed under the Apache License, Version 2.0 — http://www.apache.org/licenses/LICENSE-2.0

Citation

@misc{aileycore12b_2026,
  title  = {AileyCore-12B: A Gemma 4 fine-tune for the A!ley assistant},
  author = {van de Loo, Simon and OpenM!nded},
  year   = {2026},
  note   = {Round 3: fresh QAT fine-tune, native reasoning and tool formats, identity without a system prompt}
}