Maestro1-9B
A 9B multimodal reasoning model — math, code, and deep thinking that can see.
Vection Labs
Weights · Benchmarks · Quickstart · Limitations
Abstract
Maestro1-9B is a dense, 9-billion-parameter vision-language model built for hard problems: multi-step mathematical proof, competitive-programming-grade code synthesis, and visual reasoning over images and video — within a single model and a single context window of up to 1M tokens.
It is designed for users who care less about chat pleasantries and more about whether the model can actually solve the thing: derive the bound, find the bug, read the diagram, finish the proof. Maestro1-9B pairs an explicit step-by-step reasoning mode with native multimodal perception, so the same chain of thought that solves a math olympiad problem can also reason about a chart, a UI screenshot, or a short clip.
Highlights
- Reasoning-first. Produces structured, inspectable chains of thought for math, logic, and code.
- Genuinely multimodal. Images and video are first-class inputs, not bolted-on captioning.
- Long context. Up to 1M tokens via interleaved multimodal RoPE — whole codebases, long papers, or long videos in a single prompt.
- Open weights. Apache-2.0,
transformers-native, single-file deployment. - 9B dense. Runs on a single modern accelerator; no mixture-of-experts routing to manage.
Model overview
| Parameters | 9B (dense) |
| Modalities | text, image, video → text |
| Context window | up to 1,000,000 tokens (interleaved multimodal RoPE) |
| Precision | bfloat16 |
| Architecture | decoder-only transformer LM + native vision encoder |
| License | Apache-2.0 |
| Library | 🤗 transformers (AutoModelForImageTextToText) |
Intended use
Maestro1-9B targets technical assistance and research:
- Step-by-step math and quantitative reasoning.
- Code generation, explanation, debugging, and review.
- Visual question answering and document/diagram/chart understanding.
- Video understanding over short clips.
- Long-document and long-context analysis.
It is not intended for high-stakes decisions without human review, nor as a source of truth for medical, legal, or financial advice.
Benchmarks
Reasoning, math & code
| Benchmark | Setting | Maestro1-9B |
|---|---|---|
| GSM8K | 0-shot CoT, exact match | — |
| MATH-500 | 0-shot CoT, exact match | — |
| AIME 2024 | 0-shot, pass@1 | — |
| HumanEval | 0-shot, pass@1 | — |
| MBPP | 3-shot, pass@1 | — |
| MMLU | 0-shot | — |
Multimodal
| Benchmark | Setting | Maestro1-9B |
|---|---|---|
| MMMU (val) | 0-shot | — |
| MathVista (testmini) | 0-shot | — |
| DocVQA (val) | 0-shot, ANLS | — |
Quickstart
from transformers import AutoModelForImageTextToText, AutoProcessor
import torch
model_id = "vectionlabs/Maestro1-9B"
proc = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="auto"
)
messages = [{
"role": "user",
"content": [
{"type": "image", "image": "https://example.com/diagram.png"},
{"type": "text", "text": "Explain what this diagram proves, step by step."},
],
}]
text = proc.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
from qwen_vl_utils import process_vision_info
imgs, vids = process_vision_info(messages)
inputs = proc(text=[text], images=imgs, videos=vids, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024)
print(proc.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])
Text-only works the same way with a plain {"type": "text", ...} message.
Prompting tips
- For math/logic, ask the model to reason step by step; it is tuned to externalize its work.
- For code, specify language, constraints ("no external libraries"), and the exact I/O contract.
- For vision, put the image/video before the question in the message content.
- Lower temperature (0.2–0.7) for deterministic reasoning; raise it for brainstorming.
Deployment
- GPU: a single 24–80 GB GPU in bf16 (
device_map="auto"). - Serving: compatible with standard
transformersgeneration; for high throughput use a vision-capable serving stack.
Limitations & responsible use
- Maestro1-9B can be confidently wrong. Verify mathematical and factual claims.
- Generated code may be insecure or incorrect — review before running, never execute untrusted output.
- Long-context and long-video inputs increase latency and memory substantially.
- It inherits the biases and failure modes of large web-trained models. Do not use it for surveillance, manipulation, or any use that violates applicable law or the Apache-2.0 terms.
- No audio modality.
Citation
@misc{vectionlabs2026maestro1,
title = {Maestro1-9B: A Multimodal Reasoning Model},
author = {Vection Labs},
year = {2026},
url = {https://huggingface.co/vectionlabs/Maestro1-9B}
}