Name Generator (GGUF)

GGUF quants of aimeri/name-generator, a 3.4B fine-tune of Ministral 3. Ask it for a name and it replies with one name and nothing else, drawn from real naming data across 333 countries and historical states.

v2 is out: aimeri/name-generator-v2-GGUF also invents fantasy names (elves, dwarves, orcs, dragons, fantasy kingdoms) and keeps the real ones.

File	Quant	Size	Same top token as 16-bit	Mean KL divergence
`name-generator-Q8_0.gguf`	Q8_0	3.65 GB	98.8%	0.0018
`name-generator-Q6_K.gguf`	Q6_K	2.82 GB	97.1%	0.0055
`name-generator-Q5_K_M.gguf`	Q5_K_M	2.47 GB	96.2%	0.0120
`name-generator-Q4_K_M.gguf`	Q4_K_M	2.15 GB	94.5%	0.0226

Q8_0 is effectively the full-precision model and needs about 4 GB of memory, so start there. Q6_K and Q5_K_M stay close to it; Q4_K_M drifts the most but still answers with plausible names of the requested origin.

The last two columns compare each file with the BF16 conversion on 130 request-and-name pairs in the chat format, about 1,700 scored tokens (llama.cpp's llama-perplexity --kl-divergence): how often the most likely next token is the same, and how far the next-token distribution drifts (lower is closer).

All files carry the ChatML chat template. <|im_end|> is the eos token, and llama.cpp adds the <s> the model was trained with. Converted from the safetensors in the main repo with llama.cpp (commit 338fce6).

Usage

llama.cpp server

llama-server -hf aimeri/name-generator-GGUF:Q8_0 --jinja -c 2048

curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "messages": [{"role": "user", "content": "What is a good male Brazilian name?"}],
  "temperature": 0.8,
  "max_tokens": 24
}'

Any llama.cpp-based app (LM Studio, Jan and others) works the same way. Send one user message per name and leave the system prompt empty, since the model was trained without one.

Settings

Temperature: 0.7–1.0. At 0 it returns the single most typical name every time.
Max tokens: around 24, which is plenty for a full name.
Context: 2048 is more than enough; requests are a sentence long.

Examples

From name-generator-Q8_0.gguf: the greedy answer, then the first four distinct samples at temperature 0.8.

Request	Greedy	Sampled
What is a good male Brazilian name?	Lucas Silva	Alexandre Silva · Felipe Barbosa · Lucas Fernandes · Filipe Nogueira
Japanese woman's name	Kobayashi Masako	Maeda Yumi · Mizukami Masako · Kamimura Natsuko · Terada Shiori
German man born in the 1950s	Michael Schmidt	Gerhard Schülke · Jürgen Beyer · Michael Engel · Hermann Hoffmann
Give me a Russian man's full name	Vladimir Sergeyevich Kolesnikov	Maxwell Sergeevich Semyonov · Sergey Dmitriyevich Lebedev · Vasily Aleksandrovich Kolesnikov · Pavel Gusev
I need a name for a Mexican grandmother	María Teresa Martínez	Olivia García · María Díaz · María Teresa Díaz · Marina Rueda
A name from the Song dynasty	Zhu Xiu	Wei Xuan · Zhong Yuan · Zhao Xiu · Qian Zhong

On 11 test requests, Q6_K, Q5_K_M and Q4_K_M gave the same greedy answer as Q8_0 for 9, 7 and 8 of them. Most differences were a different surname or patronymic ("Vladimir Nikolayevich Kolesnikov" instead of "Vladimir Sergeyevich Kolesnikov").

Prompting tips, training details and limitations are in the main model card. In short: it covers real-world and historical names only (not fantasy), it always answers with a full name, and it can occasionally miss the requested sex when sampling.