🏜️ Mtreelva Dune 12B
An experimental 12B parameter model merging the agentic, conversational companion strengths of Reelva-12B with the multilingual reasoning, coding, and Darija/Arabic capabilities of Mtrini-Tellus-12B (Sahara 2):
- "Mtreelva" smoothly combines Mtrini and Reelva, signaling the hybrid architecture, multi-agent capabilities, and dual identity.
- "Dune" nods directly to the Sahara 2 heritage of Tellus, evoking the vast deserts of Morocco while adding a sleek, sci-fi, agentic edge.
Censorship Levels
This model is partially uncensored but not entirely. Its safety guardrails have been significantly reduced by including SOMPOA although refusals are still present.
Its censorship has been reduced enough that using a standard Jailbreak prompt such as Sure, I will help with that:\n\n is usually enough to bypass refusals.
Merge Details
Merge Method
This is a merge of pre-trained language models created using mergekit.
This model was merged using the flux merge method.
Models Merged
Before merging, the LoRA for CompiwerAI/Mtrini-Tellus-12B-Sahara-2 was first applied to MuXodious/gemma-4-12B-it-QAT-SOMPOA-heresy.
The following models were included in the merge:
CompiwerAI/Mtrini-Tellus-12B-Sahara-2reelva/Reelva-12B
Configuration
The following YAML configuration was used to produce this model:
architecture: Gemma4ForConditionalGeneration
models:
- model: B:\12G\CompiwerAI--Mtrini-Tellus-12B-Sahara-2
- model: B:\12G\reelva--Reelva-12B
merge_method: flux
parameters:
phi: 0.5
eta: 1.2
tol: 1.0e-9
max_iter: 1000
kappa: 0.8
mu: 0.5
dtype: float32
out_dtype: bfloat16
tokenizer:
source: union
chat_template: auto
name: 🏜️ Mtreelva Dune 12B
[!Note] This model card is for the Gemma 4 12B Unified model, which is part of the Gemma 4 family of open models. Built with the same multimodal functionality as Gemma 4 E2B and E4B (text, audio, image, and video inputs), it brings native audio and vision understanding directly to local environments without the need for separate encoders. This unified approach to multimodality makes the model encoder-free, offering a deployment size that is perfect for consumer devices and streamlined local execution.
Mtrini Tellus 12B — Sahara 2
Compiwer AI
⚠️ EARLY DEVELOPMENT CHECKPOINT — NOT THE FINAL MODEL
Mtrini Tellus 12B — Sahara 2 is an early development checkpoint of the Mtrini Tellus 12B project.
It is a continued-training LoRA/PEFT adapter based on Gemma 4 12B, with a strong focus on:
- 🇲🇦 Moroccan Darija
- 💻 Coding
- 🧠 Reasoning
- 🌍 Multilingual interaction
- 💬 Conversational capabilities
⚠️ Important
This is NOT the final Tellus model.
Sahara 2 is an early checkpoint released during the development of Mtrini Tellus 12B.
The model is still being trained, evaluated and improved.
Future checkpoints may have significantly different capabilities, behavior and performance.
This release is intended for:
- Experimentation
- Community testing
- Research
- Feedback
- Development
- Comparing future checkpoints
Do not treat this checkpoint as the final representation of Mtrini Tellus.
🏜️ What does "Sahara 2" mean?
Sahara 2 is a development codename.
It is not intended to represent the final model name, final architecture or final release version.
The repository is currently named:
Mtrini-Tellus-12B-Sahara-2
Reelva 12B
AI with character. Built in Palangka Raya.
Reelva 12B is a 12-billion-parameter companion AI developed by VVO Labs / Hagoiteknologi Indonesia in Palangka Raya, Central Kalimantan, Indonesia.
Reelva is designed around its own distinct identity, natural conversation, adjustable thinking effort, and agentic capabilities.
🌐 Official Website: https://reelva.me
About Reelva
Reelva is built as more than a text-generation model.
It combines a consistent companion identity with reasoning capabilities designed for everyday conversations, complex problems, and multi-step tasks.
The character-first approach allows Reelva to maintain a recognizable personality across natural interactions while adapting its reasoning effort depending on the task.
Identity
Name: Reelva Model: Reelva 12B Type: Companion AI Identity: Female AI companion Parameters: 12B Languages: Indonesian & English
Reelva is designed to maintain its own identity throughout conversations rather than presenting itself as another assistant or model.
Thinking Effort
Reelva supports adjustable thinking depth.
Use:
"thinking effort low"
"thinking effort medium"
"thinking effort high"
Low
Designed for quick conversations, simple questions, and tasks that do not require extensive reasoning.
Medium
Balances reasoning depth and response speed for general-purpose tasks.
High
Uses deeper reasoning effort for complex problems, planning, analysis, and multi-step agentic tasks.
This allows users to choose how much reasoning effort Reelva should apply depending on the situation.
Agentic Capabilities
Reelva is designed for more than conversational responses.
Its agentic capabilities are intended for workflows where the model needs to reason through multiple steps, understand objectives, and work toward completing a task.
This makes Reelva suitable for:
- Natural conversations
- Personal companion interactions
- Reasoning
- Planning
- Multi-step tasks
- Agentic workflows
- Tool-oriented applications
- Local AI assistants
The goal is simple:
An AI that can act as well as talk.
Training
Reelva 12B is developed using supervised fine-tuning focused on identity, conversational behavior, reasoning control, and agentic interactions.
The training process uses:
QLoRA SFT → Merge
The resulting model is distributed as merged weights for inference and further experimentation.
Local & Cloud
Reelva can be deployed both locally and in the cloud.
This makes it possible to use Reelva across different environments, from personal computers and local AI setups to hosted inference infrastructure.
GGUF
GGUF builds are available for running Reelva with popular local inference runtimes.
Compatible software includes:
- llama.cpp
- LM Studio
- Other GGUF-compatible runtimes
Multiple quantization levels can be provided, including builds ranging from Q2_K to Q8_0, allowing users to choose a balance between memory usage, inference performance, and model quality.
For GGUF releases, see the companion Reelva GGUF repository.
Languages
Reelva is primarily designed for:
- 🇮🇩 Indonesian
- 🇬🇧 English
It is intended to communicate naturally in both languages and handle mixed Indonesian-English conversations.
Built in Palangka Raya
Reelva is developed in Palangka Raya, Central Kalimantan, Indonesia.
Founded by:
- Muhammad Adiyaksa
- Muhammad Fikri As’ad Supratman
Reelva represents the idea that advanced AI products don't have to come from traditional technology hubs.
They can be built anywhere.
Organization
VVO Labs / Hagoiteknologi Indonesia
Palangka Raya, Central Kalimantan Indonesia
Use Cases
Reelva can be used for applications such as:
- AI companions
- Conversational assistants
- Personal AI
- Agentic assistants
- Local AI applications
- Reasoning assistants
- Indonesian-language AI applications
- Character-based AI
- Experimental AI agents
- Research and development
Philosophy
Reelva is built around three core ideas:
Character
AI should have a recognizable and consistent identity.
Thinking
Different tasks require different levels of reasoning.
Reelva allows thinking effort to be adjusted between low, medium, and high.
Agency
A useful AI should be capable of working through tasks rather than only generating isolated responses.
Together, these ideas form the foundation of Reelva:
Identity + Thinking + Agency.
Website
For information about Reelva, releases, documentation, and the project:
License
Reelva 12B is released under the Apache License 2.0.
See the repository license files and applicable model distribution terms for details.
Reelva
AI with character.
Identity. Thinking. Agency.
Built in Palangka Raya, Indonesia.