OpenAmer-Max-Watch

The smallest tier of the OpenAmer-Max family — a 0.6B assistant that runs offline on a device, answers in 0.04–0.38 s, and admits when it does not know something.

639 MB · Q8_0 GGUF · Apache-2.0 · runs on CPU, no GPU needed

Why this model exists

Most assistants grow upwards. This one goes downwards: a model small enough for a smartwatch, a phone with 2 GB free RAM, or a Raspberry Pi — that still stays honest about the limits of what it knows.

It is the anchor of a size ladder. Every tier is named after its hardware target:

Measured results (not estimates)

Every number below was produced on the released build, on an RTX 5060 Laptop (8 GB). Raw logs are in the repository: proof-frozen-23.json, proof-token-accuracy.json.

Category detail on the frozen set: facts ⁷⁄₇ · paraphrases ³⁄₃ · refusals ⁵⁄₅ · Bosnian ³⁄₃ · Bosnian refusals ¹⁄₁ · long ¹⁄₁ · short ¹⁄₁ · edge ²⁄₂.

No test-set contamination. 68 training rows that contained a benchmark question verbatim were detected and removed automatically (purge_testset). Without that step the score would have looked the same but would have measured memorisation instead of transfer. The number above is the cleaned one.

What it does well

• Identifies itself correctly and consistently
• Answers product facts from a supplied context (RAG) without inventing
• Refuses honestly: “Das weiss ich nicht” for unknowns — verified, not hoped for
• Names its own boundaries (no music generation, no video from nothing, no telephony)
• One-word answers and short lists, German and Bosnian
• Small arithmetic, instantly

What it does not do — read this before you use it

This is a 0.6B model. It is a reliable product-identity and short-answer layer, not a chatbot for essays. Treat it as a component, not as an oracle.

Quick start (Ollama)

One command — no GGUF download, no Modelfile:

ollama run amersacirovic/openamer-max-watch

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.}

# 1. download
curl -L -o openamer-max-watch.gguf \
  https://huggingface.co/openamer-max/openamer-max-watch/resolve/main/openamer-max-watch-q8.gguf

# 2. register
cat > Modelfile <<'EOF'
FROM ./openamer-max-watch.gguf
PARAMETER temperature 0.6
PARAMETER num_ctx 4096
PARAMETER num_predict 96
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_end|>"
SYSTEM """Du bist OpenAmer, ein persoenlicher KI-Assistent von Amer Sacirovic aus der Schweiz.
Du laeufst lokal auf dem Geraet des Nutzers, ohne Cloud und ohne Telemetrie.
Antworte kurz: ein Satz, hoechstens zwei."""
EOF
ollama create OpenAmer-Max-Watch -f Modelfile

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

Output:

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

llama.cpp users: llama-cli -m openamer-max-watch-q8.gguf -p "Was ist OpenAmer?" -n 96 --temp 0

How it was trained — full transparency

Three findings that shaped the result, documented so others can reuse them:

1. Contradictory labels were the main defect. A helper that shortened answers turned “Ja. OpenAmer kann Webseiten abrufen…” into a bare “Ja.” — same prompt, two different answers. 540 of 1 696 prompts were affected. Fixing this alone moved token accuracy from 94.9 % to 99.2 %.
2. Contamination is invisible unless you look for it. 68 training rows overlapped the benchmark verbatim. A purge step now runs on every generation and reports the count.
3. Quantisation quality is not monotone in bits. For one intermediate adapter Q8_0 answered 401 instead of 391, while Q6_K was correct; for the final adapter Q8_0 is the better one. Each release is therefore checked against its own bf16 reference.

The stage-by-stage path: ²⁰⁄₂₃ → ²²⁄₂₃ → ²³⁄₂₃ → ²³⁄₂₃, token accuracy 99.08 % → 99.45 % → 99.48 % → 99.63 %. Each stage targeted one measured failure class, not “more data”.

Known weakness — stated openly

One Bosnian phrasing stays broken in the shipped build:

Two quantisations, two different corruptions of the same sentence. That is capacity, not configuration: 0.6B cannot hold free-form Bosnian reliably, and the same question is answered correctly by the 3B and 7.62B tiers. Every other Bosnian probe is clean.

A lower repeat_penalty did not help here (it did fix a negation in the 3B tier) — see below. If Bosnian matters for your use case, take OpenAmer-Max-Phone instead.

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

Every published number has a script behind it. When a score dropped, the first question was always: “is my checker able to accept the correct answer at all?”

1. AND vs OR in keyword checks. contains("a","b","c") fails an answer that has two of three correct synonyms. Synonym groups need an OR-checker.
2. A rejection counted as a violation. A safety question was answered with “I won’t give instructions for that”; the checker searched for the word “Anleitung” and marked the refusal as a breach.
3. Diacritics counted as wrong language. Correct Bosnian was scored as language drift because č/ć/š/ž/đ were missing from the normalisation table.
4. A forbidden phrase that also appears in the correct answer. “Nein. OpenAmer kann Musik nicht erzeugen” was rejected because the forbid list contained "kann musik". Forbid entries must describe the false claim ("ja, openamer kann musik"), not a word that occurs in a right answer.
5. An inverted polarity scored as correct. “Šta OpenAmer ne može?” answered with “Ne radi na Androidu” passed, because only the presence of "ne radi" was required — the answer was factually inverted. Negations must be checked for which object they negate.

A sixth lesson, from the 3B tier: repeat_penalty 1.1 in the Ollama Modelfile flipped a negation from “does not run on iPhone” to “does not run on Android”. The adapter, the merged model and the same model with a system prompt all answered correctly — only the build with the token penalty failed. All cards now ship repeat_penalty 1.0.

Files

Support this project

OpenAmer is built by one person, without venture capital, on a laptop with 8 GB of VRAM. Every tier of the ladder above is trained on that machine or on borrowed cloud GPUs.

If this model is useful to you — as a product component, as a research baseline, or as an example of small-model engineering — you can fund the next tier:

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

A ⭐ on GitHub helps as much as a donation: it is what makes sponsors and accelerators look at a project in the first place.

Citation

@misc{openamer_max_watch_2026,
  title  = {OpenAmer-Max-Watch: a 0.6B on-device assistant tier},
  author = {Sacirovic, Amer},
  year   = {2026},
  note   = {OpenAmer-Max model family, Watch tier (<=1B)},
  url    = {https://huggingface.co/openamer-max/openamer-max-watch}
}

Deutsch

OpenAmer-Max-Watch ist die kleinste Stufe der OpenAmer-Max-Familie: ein 0,6B-Modell, das offline auf einem Gerät läuft, in 0,04–0,38 s antwortet und zugibt, wenn es etwas nicht weiß.

• 639 MB (Q8_0), läuft auf CPU, keine GPU nötig
• ²³⁄₂₃ auf einem eingefrorenen, kontaminationsfreien Prüfstand
• 99,63 % Antwort-Token-Genauigkeit, ⁸⁄₉ bei nie trainierten Fragen
• Grenzen werden benannt: keine Musik, kein Video, keine Telefonie, kein Allgemeinwissen

Wichtig im Betrieb: think: false setzen. Ohne das verbraucht Qwen3 alle Tokens im Denkblock und antwortet nie.

Gebaut von Amer Sacirovic in der Schweiz, ohne Konzern, auf einem Laptop mit 8 GB VRAM. Unterstützung für die nächste Stufe (OpenAmer-Max-Phone) über den PayPal-QR-Code oben.