Focus-5-GGUF

Focus 5

A 495 MB GGUF that thinks before it answers, replies in your language, and reaches for a tool instead of guessing.

Focus-5.gguf     495,111,584 bytes · 472.18 MiB · 310 tensors · 40,960 ctx
sha256           4190e0525083ad020e6f1b2a23a2c14497c0eba1a6cc047d7fc1e01fa5c7faf9

---

Read this first

Focus 5 is not trained from scratch, and it is not a frontier model.

The weights are Qwen3-0.6B at Q6_K, copied byte for byte and not retrained — all 310 tensors verify identical to the base file. What this build changes is the metadata block at the head of the GGUF: the identity, the tokenizer.chat_template, and a set of focus.* / sinythos.* keys that a runtime reads back to configure itself.

That is the honest description, and it is written into general.description inside the file itself so nobody downstream can mistake it for something else.

What 500 MB can and cannot be

500 MB of Q6_K weights is about 0.6 billion parameters. It can be genuinely multilingual, open a reasoning block on every turn, call tools correctly, and write working code for ordinary tasks. It cannot match a frontier model on benchmarks — not with better training, not with a better dataset, not with any quantisation trick. The gap is information-theoretic, not effort-based.

What does close a large part of the practical gap is the scaffolding: a small model that searches the web, scrapes the page, and computes with an exact calculator beats a much larger model working from memory alone on most real tasks. The six laws below exist to push it towards those tools on every turn where they apply.

---

The six laws

Baked into tokenizer.chat_template, so they apply even when the caller sets no system prompt at all. A user-supplied system prompt is kept, but framed as operator instructions that apply within the laws.

1. Think first, in your language. Opens <think> on every non-trivial turn. Detects the user's language and answers in it.
2. Do not guess — look it up. Anything recent, changing, or uncertain goes to web_search + scrape_page before answering. Inventing a fact, date, quote or URL is the one unforgivable failure.
3. Never do arithmetic in your head. Every calculation goes to calculate.
4. Code is complete or it is not delivered. Real imports, real error handling, no TODO, no ....
5. 3D is real geometry. Real mesh, PBR materials, shadows — not coloured boxes.
6. Break your own work before shipping it. Adversarial second pass, then say what was fixed.

The directive is readable straight out of the file under the sinythos.system_directive metadata key.

---

Run it

llama-server -m Focus-5.gguf -c 40960 --jinja --host 0.0.0.0 --port 8080

--jinja is not optional. Without it llama.cpp ignores the embedded chat template, the six laws never reach the model, and you get a stock Qwen3-0.6B. That is the single easiest mistake to make with this file, and the symptom is "it behaves like plain Qwen".

LM Studio and Jan read the template from the file on their own, so no flag is needed there — just drag the file in.

Recommended sampling (Qwen3 thinking mode): temperature 0.6, top_p 0.95, top_k 20, min_p 0. Do not use greedy decoding — it makes this family loop.

Tools it expects

web_search, scrape_page, calculate, run_code, write_file, make_asset — listed in the sinythos.tools.required metadata key, advertised in Qwen's XML-wrapped-JSON tool-call format. A reference runtime that implements all six, including an exact BigInt-rational calculator, is linked below.

---

Verify it yourself

sha256sum Focus-5.gguf
# 4190e0525083ad020e6f1b2a23a2c14497c0eba1a6cc047d7fc1e01fa5c7faf9

The build is reproducible: fetch Qwen3-0.6B-Q6_K.gguf from unsloth/Qwen3-0.6B-GGUF, re-mux it with this project's build script, and you get the same digest.

---

Licence

Apache 2.0, inherited from Qwen3-0.6B — the weights are Qwen's work and are unmodified. The chat template, the system directive and the metadata are this project's, also Apache 2.0.