GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
vectionlabs
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
vectionlabs
Model Tree
Base
this model
Input Modalities
Text
Image
Video
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Maestro 2 is a 9B dense multimodal model built for engineering work you can run on one GPU.
It thinks natively inside <think> blocks, writes and debugs real code, designs frontends that do
not look machine-generated, and reads images, screenshots, diagrams and video as first-class input.
Context window: 262,144 tokens.
It is the successor to Maestro 1 — the most-liked model we have shipped — and it keeps that model idea intact: one small brain that reasons, codes, and sees, rather than three specialists you have to switch between.
<think> reasoning on by default; switch it off with
enable_thinking=False when you want an instant answer.| Parameters | 9.65B dense |
| Modalities | text, image, video -> text |
| Context window | 262,144 tokens native(In version 2.1, the context will be updated to 1M.) |
| Precision | bfloat16 |
| Architecture | Qwen3.5 (dense) + native vision encoder |
| License | Apache-2.0 |
| Library | transformers (AutoModelForImageTextToText) |
Larger MoE siblings: Maestro-2-9B-Preview (35B-A3B) and Salience-1.5-Flash (30B-A3B).
Thinking is on by default — the model reasons inside <think>...</think> before answering, and
serving stacks surface it as reasoning_content. Pass enable_thinking=False to
apply_chat_template for direct answers.
Maestro 2 emits XML-style tool calls (<tool_call><function=...><parameter=...>), parsed
natively by the vLLM and SGLang tool parsers for this model family. Provide schemas through the
chat template tools argument.
python
from transformers import AutoModelForImageTextToText, AutoProcessorrepo = "vectionlabs/Maestro-2-9B-Preview"proc = AutoProcessor.from_pretrained(repo)model = AutoModelForImageTextToText.from_pretrained(repo, dtype="auto", device_map="auto")messages = [{"role": "user","content": [{"type": "text", "text": "Build a responsive pricing card in React + Tailwind."}],}]text = proc.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = proc(text=[text], return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=2048)print(proc.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])
Requires a recent transformers (>= 5.8). Images work the same way with
{"type": "image", "image": ...} content items.
Dense 9B quantizes gracefully, unlike the sparse-MoE tiers of this family:
| quant | size | fits |
|---|---|---|
| Q8_0 | ~10 GB | 12 GB card, best quality |
| Q5_K_M | ~6.5 GB | 8 GB card [recommended] |
| Q4_K_M | ~5.6 GB | 6 GB card, still coherent |
viewBox and a stated grid; it plans geometry before drawing.temperature=0.7, top_p=0.95, top_k=20, presence_penalty=1.1. Lower the
temperature to ~0.3 for precise patches.bibtex
@misc{vectionlabs2026maestro2,title = {Maestro 2 Preview: A 9B Multimodal Engineering Model},author = {Vection Labs},year = {2026},url = {https://huggingface.co/vectionlabs/Maestro-2-9B-Preview}}