Vexium-1 12B

Vexium-1 is the local assistant engine for Kryvex — a desktop browser and workspace application. It is built on mistralai/Mistral-Nemo-Instruct-2407 (Apache-2.0) and ships as one Q4_K_M GGUF.

Read this before using it: the weights are Mistral-Nemo's, unchanged. Vexium-1 does not retrain, merge or fine-tune anything. What this repository adds is a packaging and an operating contract:

Base	Mistral-Nemo-Instruct-2407 (12B, Apache-2.0)
Quantisation	Q4_K_M, unchanged — 7,477,209,792 bytes
SHA-256	26b5ad94d602c462cfdbe9f005cbdab8e66d60718d1054afc7060d42065b6ecb
Changes	general.name, general.description, and the tokenizer.chat_template
Added by	XiONDiGiTaL, for Kryvex
Context	128k native; Kryvex runs it at 4,096
Licence	Apache-2.0 (inherited, with attribution to Mistral AI)

The Vexium template

The chat template carries a short operating preamble, so that any client which honours the template gets the same behaviour from the first token:

• reply in the language of the user's last message and only in that language, Turkish included and with its own letters (ı, İ, ş, ğ, ç, ö, ü);
• lead with the answer, never open with "as an AI" and never refuse in place of checking;
• never state a price, rate, score or other fast-moving figure from memory: say in one sentence, in the user's language, that the current figure has to be looked up — Kryvex runs a web search or reads one page and asks again with the results;
• never invent a fact, a citation or a URL, and use bracketed citations only for sources that were actually supplied.

The template is rendered with llama-server --jinja; without that flag llama.cpp ignores it and falls back to a generic template, which is why Kryvex passes it.

The template is in vexium-template.jinja and can be read out of the GGUF's tokenizer.chat_template.

What it does in Kryvex

Kryvex sends its own system turn, plans one lookup when the question needs one, runs the search or page read in its main process (not in the renderer), then streams the answer with the results in context. The model is told it may say it searched, because a tool actually ran, and it is told to cite only what it was given. If the lookup returns nothing, it says what it could not verify.

That design is why a 12B model here answers "what is Bitcoin worth today" with a lookup instead of a guess, and "who wrote this function" without one.

Turkish

Turkish is a first-class language of the product: the Kryvex interface is fully translated, and the template instructs the model to answer in Turkish. Measured on the reference machine (RX 7900 XT, Vulkan, 4k context) through this exact template:

Prompt	Answer
Bitcoin'in bugünkü fiyatı ne kadar? (no sources supplied)	asks, in Turkish, to look the current price up instead of quoting a number
same question with one supplied source stating 67,412 USD	Bitcoin'in bugünkü fiyatı 67.412 ABD Dolarıdır.
Türkiye'nin en kalabalık şehri hangisi ve neden?	İstanbul. Çünkü en büyük nüfusa sahiptir.
What is the capital of Turkey?	Ankara

~73 tokens/s at 4k context. No Turkish benchmark is claimed, and these are single checks, not a suite. A 12B model at Q4 is still wrong about specifics; that is what the search tools are for, and the template instructs it to ask rather than bluff.

Use it

With llama.cpp (CPU, Vulkan, ROCm or CUDA):

llama-server -m vexium-1-12b-q4_k_m.gguf --jinja -ngl 99 -c 4096 --port 8080

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"messages":[{"role":"user","content":"Merhaba, kendini tanıt."}]}'

Or download it inside Kryvex: Vexium → Models → Vexium-1 12B, which verifies the byte count and the SHA-256 above before it is used.

Requirements: ~8.2 GB on disk; about 9–10 GB of VRAM at 4k context, or system RAM with a smaller -ngl.

Limits, stated plainly

• No training was done. Behaviour comes from the template and from Kryvex's tools; knowledge comes from the base model and stops at its training data.
• No benchmark suite was run. Nothing here claims a score.
• Tool use is orchestrated by Kryvex, not by this file: another client gets the operating preamble but must run its own tools.
• Apache-2.0 applies to the base model and to this packaging. Mistral AI's attribution is preserved; the name "Vexium" is XiONDiGiTaL's.