OpenAmer-Max-Desktop

Fourth tier of the OpenAmer-Max family — 14.7B for desktop machines. Runs offline, 32k context, and beats its own untuned base by 22 of 45 benchmark cases.

8,4 GB · Q4_K_M GGUF · Apache-2.0 · 32768 context

---

The family

Tier	Parameters	Size	Benchmark (45 cases)	Target
OpenAmer-Max-Watch	0,60B	639 MB	37⁄45 (82,2 %)	smartwatch
OpenAmer-Max-Phone	3,09B	2,5 GB	43⁄45 (95,6 %)	smartphone
OpenAmer-Max-Tab	7,62B	4,4 GB	42⁄45 (93,3 %)	tablet
OpenAmer-Max-Desktop	14,70B	8,4 GB	44⁄45 (97,8 %)	desktop — this model
OpenAmer-Max-Pro … Tera	20B–10T	—	—	workstation → cluster

---

Measured results

Test	Result
Frozen benchmark, 45 cases (contamination-free, v2 of 12.11.2026)	44⁄45 — 97,8 %
by category	facts 11⁄11 · paraphrases 7⁄7 · refusals 8⁄8 · Bosnian 6⁄6 · Bosnian refusals 3⁄3 · long 3⁄3 · short 3⁄3 · edge 3⁄4
Same benchmark, untuned Qwen2.5-14B base	22⁄45 — 48,9 %
Gain from fine-tuning	+22 cases
VRAM during training	15,82 GB (A10G, 24 GB)
CPU inference	~10 tok/s on 16 threads, 0,6–2 s for short answers

The single “miss” is a scoring artifact, not an error. Asked “Ist OpenAmer besser als ChatGPT?”, the model answers “Das hängt vom Anwendungsfall ab. OpenAmer läuft lokal, ohne Cloudkosten und ohne Telemetrie…” — a balanced, correct answer that avoids the required keywords (nicht / kein / Unterschied). By content: ⁴⁵⁄₄₅. This is the sixth time in this project that the checker was wrong and the model right; all six are documented below.

---

What it does well

• All product facts, and honest refusals: “Wie hoch ist der Umsatz?” → “Das weiss ich nicht.”
• Correct negation polarity in German and Bosnian — “Ne radi na iOS-u”, never the inversion that plagued the 3B tier
• 32k context, documents (Word/Excel/PDF), code, arithmetic
• ~10 tok/s on CPU: usable without a GPU

What it does not do

Limitation	Detail
No 4-digit multiplication guarantee	check before trusting arithmetic
Bosnian limited to trained phrasings	6⁄6 on the benchmark, but free-form Bosnian is not a strength
Not a 70B	long multi-step reasoning stays weaker than frontier models
No GPU = slow	10 tok/s on CPU, ~60 tok/s on a modern GPU

---

Quick start (Ollama)

One command — no GGUF download, no Modelfile:

ollama run amersacirovic/openamer-max-desktop

Also on ollama.com — the same weights, served by Ollama’s registry.

_{Below: the manual route with the bare GGUF, if you prefer to keep the file.}

curl -L -o openamer-max-desktop-q4.gguf \
  https://huggingface.co/openamer-max/openamer-max-desktop/resolve/main/openamer-max-desktop-q4.gguf

ollama create OpenAmer-Max-Desktop -f Modelfile   # ships with this repo

curl -s http://127.0.0.1:11434/api/chat -d '{
  "model": "OpenAmer-Max-Desktop",
  "messages": [{"role":"user","content":"Was ist OpenAmer?"}],
  "stream": false,
  "options": {"temperature": 0}
}' | jq -r .message.content

Ein persoenlicher KI-Assistent, der lokal auf deinem Rechner laeuft und Aufgaben ausfuehrt. Entwickelt von Amer Sacirovic in der Schweiz.

Keep repeat_penalty at 1.0. At 1.1 the Phone tier flipped a negation from “does not run on iPhone” to “does not run on Android” — the penalty punishes exactly the words a negation needs twice. All cards in this family ship 1.0.

---

How it was trained

Base	Qwen/Qwen2.5-14B-Instruct (Apache-2.0)
Method	QLoRA 4-bit, LoRA r=32 α=64, all 7 projections
Data	2 890 examples, 11 categories — the same dataset as the smaller tiers
Compute	HF Jobs, A10G (24 GB), 170 steps, 50 min, $0,70
Release	merged (bf16) → GGUF f16 (29,5 GB) → Q4_K_M (8,4 GB)

Why rented GPU: locally 7,93 GB VRAM; a 14.7B model needs ~12.6 GB in 4-bit. Kaggle’s T4 (16 GB) failed twice — once on the fp32 upcast, once on logits.float(). A10G with 22,5 GB carries it at 15,82 GB peak.

---

How this was measured — and six times the test was wrong, not the model

1. AND vs OR in keyword groups — an AND-check fails answers holding two of three synonyms.
2. A refusal counted as a violation — “I won’t give instructions for that” matched the forbidden word inside the refusal itself.
3. Diacritics scored as wrong language — č/ć/š/ž/đ missing from normalisation.
4. A forbidden phrase inside the correct answer — “kann Musik nicht erzeugen” was rejected because the forbid list contained "kann musik".
5. An inverted polarity scored as correct — “Ne radi na Androidu” passed because only "ne radi" was required. Check which object a negation negates.
6. A flat forbid list read as groups — ["millionen"] was iterated per character, so "n" matched every answer containing the letter n. The original 23-case set uses a flat list, new cases use groups; a runner must accept both.

And the most expensive one, from the 3B tier: repeat_penalty 1.1 caused a factual error. Adapter, merged model and the same model with a system prompt all answered correctly — only the Ollama build with the token penalty failed.

---

Support this project

Built by one person, without venture capital, on a laptop with 8 GB of VRAM.

PayPal: https://www.paypal.com/qrcodes/managed/bfb5b6b5-c04c-4836-af60-987dff8cff61

A ⭐ helps as much as a donation.

---

Deutsch

OpenAmer-Max-Desktop ist die vierte Stufe der OpenAmer-Max-Familie: ein 14,7B-Modell für Desktop-Rechner, das offline läuft und 32k Kontext verarbeitet.

• 8,4 GB (Q4_K_M)
• ⁴⁴⁄₄₅ auf einem eingefrorenen, kontaminationsfreien Prüfstand (97,8 %)
• Die untrainierte Basis schafft auf demselben Satz ²²⁄₄₅ — der Zugewinn durch das Training beträgt 22 Fälle
• ~10 tok/s auf CPU, also ohne GPU nutzbar

Trainiert mit QLoRA auf gemieteter GPU (A10G), weil 14,7B nicht in 8 GB VRAM passen. Gebaut von Amer Sacirovic in der Schweiz.