Mira 12B: talks like a person, not like an assistant

# Mira 12B **A 12B chat model that talks like a person, not like an assistant.** Short when you're short, warm when you're down, straight to the point when you're busy. No bullet-point essays, no "Great question!", no em dashes. It thinks before it answers, and its thinking reads like someone mulling it over, not a to-do list. It keeps the base model's brains: math, code and tool calling stay where they were, and it runs on a single consumer GPU or in your device Mira 12B intro: a prompt, Gemma 4 12B's long markdown answer next to Mira's short reply, real replies in English, Japanese and Chinese, and the style and ability numbers from the Results table

Mira 12B intro: a prompt, Gemma 4 12B's long markdown answer next to Mira's short reply, real replies in English, Japanese and Chinese, and the style and ability numbers from the Results table

## Before and after Same prompt, same sampling. Held-out prompts the model never saw, outputs unedited. English: is it fine to dislike a movie everyone calls a classic. Mira answers in a few honest sentences; Gemma 4 12B writes a long breakdown with headings

English: is it fine to dislike a movie everyone calls a classic. Mira answers in a few honest sentences; Gemma 4 12B writes a long breakdown with headings

Japanese: worried a dating-app match might be married. Mira tells him searching won't work and to just ask her; Gemma 4 12B lists numbered points

Traditional Chinese: someone lonely at 2:30am. Mira answers in two short lines; Gemma 4 12B writes many paragraphs

English: a playful order to make five slides. Mira plays along and lists the five slides; Gemma 4 12B writes a role-play monologue

## It writes for you Rewriting your own message so it sounds like you, a poem for someone you love, a lullaby for a baby who won't sleep, a text you're too tired to word: Mira gives you one good version you can use right away, instead of three options and a lecture. Japanese: rewrite a stiff LINE message to a landlord so it sounds natural and asks about the 1,500 yen fee. Mira returns one ready-to-send message; Gemma 4 12B offers three patterns with headings

Japanese: rewrite a stiff LINE message to a landlord so it sounds natural and asks about the 1,500 yen fee. Mira returns one ready-to-send message; Gemma 4 12B offers three patterns with headings

English: a short poem for a mother's 80th birthday, warm but not a greeting card. Mira writes seven lines; Gemma 4 12B writes three options with headings

Simplified Chinese: a rhyming lullaby for an eight-month-old nicknamed Tank. Mira writes three short verses; Gemma 4 12B writes three versions with bold text and notes

English: a quick congrats text for a friend who finished his first Ironman. Mira writes two short texts; Gemma 4 12B writes a list of options

## Results Mira 12B vs Gemma 4 12B: far less markdown, dashes and assistant-style thinking, no stuck repetition, and higher GSM8K, code and tool-calling scores

Mira 12B vs Gemma 4 12B: far less markdown, dashes and assistant-style thinking, no stuck repetition, and higher GSM8K, code and tool-calling scores

Held-out prompts the model never saw, 3 samples each, default sampling. | | Mira 12B | Gemma 4 12B | |---|---|---| | Answers with markdown in casual chat | **2.6%** | 87.7% | | Answers with em/en dashes | **0.3%** | 34.5% | | Thinking that reads like an assistant plan ("the user wants…") | **4.4%** | 76.3% | | Replies in the user's language | 100% | 100% | | Answers that get stuck repeating themselves | **0%** | 0.6% | Ability kept (greedy): GSM8K **94.0%** (base 70.0%, mostly cut off mid-thought), code with unit tests **84.2%** (base 57.9%), tool calling **62.3%** (base 59.2%). ## Usage Sampling is already set in `generation_config.json` (temperature 1.0, top_p 0.95, top_k 64), so you don't need to pass anything. It also holds up at lower temperatures. ```bash vllm serve beyoru/mira-12b --reasoning-parser gemma4 --tool-call-parser gemma4 --enable-auto-tool-choice ``` ```python from transformers import AutoTokenizer, AutoModelForCausalLM tok = AutoTokenizer.from_pretrained("beyoru/mira-12b") model = AutoModelForCausalLM.from_pretrained("beyoru/mira-12b", torch_dtype="bfloat16", device_map="auto") msgs = [{"role": "user", "content": "long day. my brain is soup"}] ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=True, return_tensors="pt").to(model.device) print(tok.decode(model.generate(ids, max_new_tokens=1024, do_sample=True)[0][ids.shape[1]:])) ``` ## Limitations It can make up specifics when it doesn't know them (song titles, names, rules), so check facts before you rely on them. In about 1% of answers it closes its thinking twice and writes a second version after a stray ``; keeping only the text after the last `` fixes it. Style numbers are automatic counts, not human ratings. ## License **Non-commercial use only.** The Mira 12B weights, this card and its images are released under [CC BY-NC 4.0](LICENSE): you can use, share and adapt the model for research and personal projects with attribution, but you may not use it, or offer it or its outputs as part of a paid product or service, without written permission. For commercial licensing, contact [beyoru](https://huggingface.co/beyoru). Mira 12B is built on [Gemma 4 12B](https://huggingface.co/google/gemma-4-12b-it) by Google, licensed under the Apache License 2.0; see [NOTICE](NOTICE) and [LICENSE-APACHE-2.0](LICENSE-APACHE-2.0). ## Citation ```bibtex @misc{mira12b2026, title = {Mira 12B: a chat model that talks like a person}, author = {beyoru}, year = {2026}, howpublished = {\url{https://huggingface.co/beyoru/mira-12b}}, note = {Fine-tuned from Gemma 4 12B} } ```

Full Model Card (README)