Mentoria RV — Qwen3.6 35B A3B Preview

Mentoria RV is a Brazilian Portuguese conversational assistant specialized in explaining the Mentoria RV service and answering prospective subscribers' questions. This repository contains the merged, 4-bit MLX version of the selected supervised fine-tuning checkpoint, published as mentoria-rv-35b-a3b-preview.

O modelo foi adaptado para apresentar a Mentoria RV de forma comercial e completa, explicar seus recursos, responder sobre planos e formas de pagamento e acompanhar perguntas de continuação. Ele foi desenvolvido para usar informações confirmadas sobre o serviço, preservar o contexto da conversa e respeitar as correções do cliente.

Access to the model files requires manual approval from the repository owner. Request access on this page, wait for approval, and authenticate your downloads with the approved Hugging Face account.

Model details

The qwen3_5_moe architecture name in the MLX configuration is the compatibility implementation used for this Qwen3.6 export. It does not indicate a different base checkpoint.

Adaptation and intended use

The selected checkpoint was produced through supervised fine-tuning with an adapter, which was then merged into the base weights and converted to MLX. Training used privately maintained Mentoria RV factual material and synthetic conversational examples, including service introductions, pricing and payment questions, support boundaries, corrections, and follow-up exchanges. The training data and private run logs are not distributed here.

This release contains the selected SFT checkpoint. Subsequent reinforcement-learning candidates were evaluated separately; their weight updates are not part of this artifact. Distillation was not used for this release.

The intended use is a Mentoria RV sales and information assistant: describe the educational service, explain its features, present the 12- and 36-month plans with Pix and credit-card alternatives when asked about plans or prices, and answer follow-ups without repeatedly restarting the sales presentation.

Use the supplied mentoria_system_prompt.txt as the system message. It includes the factual context and response instructions used for deployment. Supply conversation history on subsequent turns. Review and update commercial facts before operational use; published prices and service details are a snapshot, not a live data connection. The model cannot confirm a payment, release access, or perform a cancellation without an actual tool integration.

Evaluation and its scope

The selected source adapter before MLX quantization completed a full training-data pass and passed the internal application suite twice:

These were semantic judgments by GPT-6.1-sol at xhigh reasoning effort, using a detailed, fact-backed system prompt. Thinking was enabled. The recorded evaluation used temperature 0.7, top-p 1.0, no top-k truncation, and a 2,048-token combined thinking-and-answer budget. That historical evaluation budget is not an inference limit saved in the released model.

These scores measure application instruction-following with supplied facts, not unaided factual recall or general benchmark performance. They are internal development checks used in checkpoint selection, not an independent held-out certification.

The MLX export has verified weight conversion, restored normalization vectors, conversation formatting, and functional thinking-on/off checks. It has not completed the entire 48-case semantic suite in the quantized runtime. One inspected MLX introduction omitted some Morning Call scheduling and recording details. The source checkpoint's 48/48 scores should therefore not be presented as a full validation of this quantized release. No standard public benchmark score is claimed for this artifact.

Download and run

After your access request is approved:

python -m pip install -U huggingface_hub mlx-lm
hf auth login
hf download notfakerv/mentoria-rv-35b-a3b-preview --local-dir ./mentoria-rv-35b-a3b-preview

The files are intended for Apple Silicon MLX runtimes. The five weight shards require approximately 20.19 GiB on disk; inference requires additional memory for activations, caches, and runtime overhead.

Normalization compatibility

The deployed reference used MLX 0.32.3 and MLX-LM 0.31.4.dev132 with a model-specific normalization hook. This export stores 101 normalization vectors in FP32. To reproduce its intended arithmetic, normalization and multiplication must run in FP32 and their results must be cast back to the input activation dtype.

The reviewed helper mentoria_mlx_normalization.py implements that behavior and verifies the model-specific restoration manifest. It is invoked explicitly; downloading the repository does not install or execute a runtime patch. A standard runtime that ignores this helper may behave differently. Use an MLX-LM version supporting this architecture and the supplied chat template.

For MLX-LM, run the following from the downloaded directory after reviewing the helper:

from pathlib import Path
from mlx_lm import load, stream_generate
from mlx_lm.sample_utils import make_sampler
from mentoria_mlx_normalization import apply_model_normalization

model_dir = Path.cwd()
model, tokenizer = load(str(model_dir))
assert apply_model_normalization(model, model_dir) == 101

messages = [
    {"role": "system", "content": (model_dir / "mentoria_system_prompt.txt").read_text()},
    {"role": "user", "content": "O que é a Mentoria RV?"},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=True
)
for response in stream_generate(
    model,
    tokenizer,
    prompt=prompt,
    sampler=make_sampler(temp=0.7, top_p=1.0, top_k=0),
    max_tokens=-1,  # No fixed output cap; stops at EOS or when the caller stops it.
):
    print(response.text, end="", flush=True)
print()

Set enable_thinking=False to disable thinking. The template preserves previous reasoning when it is supplied in conversation history. A client should separate thinking from the final answer when displaying replies. No fixed thinking or output-token cap is stored in generation_config.json; callers may set runtime budgets. The context window and available memory still apply.

For LM Studio, download the complete repository into a model directory such as ~/.lmstudio/models/notfakerv/mentoria-rv-35b-a3b-preview, use an MLX backend supporting the architecture, and supply the included system prompt. Thinking is controlled through the template's enable_thinking parameter. The reference deployment also applied the normalization hook above; a stock LM Studio installation does not automatically load that helper merely because it is in the model directory. This repository does not modify your LM Studio installation.

Limitations

This is a domain-specific preview and can omit facts, misunderstand short follow-ups, or produce inaccurate statements. The detailed system prompt is part of the evaluated application setup. General reasoning, coding, tool use, and long-context quality are not established by the internal application scores. Quantization and runtime differences can affect responses.

The assistant explains an educational service; it is not a personalized portfolio-management system and does not guarantee investment results. Review customer-facing outputs and keep the factual context current.

Provenance and license

The source adapter SHA-256 is 8d107eb07e9a72c23694b81ac18585b39804d84493129365b73b0a9a154032a2. Sanitized release and evaluation metadata are provided in release_metadata.json, and file hashes are listed in SHA256SUMS.

The original Qwen model is licensed under Apache 2.0. This derivative is distributed under the included LICENSE, with the modifications described in NOTICE. Manual Hugging Face approval controls access to this repository's downloads; it does not replace the license terms.