MiniMax-M3.1-Flash Joins Ultralytics Platform but Ranks 23rd of 38 in Benchmark

What MiniMax-M3.1-Flash is

MiniMax released MiniMax-M3.1-Flash on October 19, 2026. It is the fast, low-cost model in the company's M3.1 series, and MiniMax plans to release a larger Pro model later, according to Ultralytics.

The model is natively multimodal and uses a mixture-of-experts design with 428 billion parameters, 23 billion of them active per token. It has a 1-million-token context window and accepts image and video input. MiniMax built it for agentic work, coding and computer use, and says it generates more than 170 tokens per second at the median. The weights were published on Hugging Face on October 23, 2026, under the MiniMax Community License.

Listed pricing is a flat $0.15 per million input tokens and $0.60 per million output tokens.

How it works in Ultralytics Platform

MiniMax-M3.1-Flash appears under a new MiniMax provider among the class-prompted vision language models in the annotation editor. On an object detection dataset with 1 to 200 classes, it returns bounding boxes for the instances of the dataset's classes that it finds. Users review and save them like any other prediction.

• Single image: switch to Smart mode, pick the model under MiniMax and click Predict.
• Whole dataset: run Batch Annotation from the dataset page to label every unlabeled image.

Paid models run on the user's own provider API key, so a MiniMax API key must be added in the workspace's API Keys settings. Batch Annotation adds a standard Ultralytics Platform fee of $1.00 per 1,000 images, separate from what MiniMax bills. Each run saves a dataset version first, so it can be rolled back.

Ultralytics also made two settings choices after testing. The model defaults to maximum reasoning effort, which ran past the Platform's 115-second limit on more than half of the images tried. Medium scored best of low, medium and high, so the Platform uses medium. The model also places boxes more accurately in pixel coordinates than on the 0 to 1000 grid, by about 0.02 mAP50, so the Platform asks for pixels.

Benchmark results

The VLM Leaderboard uses a frozen test called UL37 v2: 10 labeled images from each of 36 detection datasets, 360 in total. The datasets cover areas such as medical imaging, aerial imagery, PCB inspection, security X-ray, retail, traffic and underwater scenes. Each image goes through the production annotation pipeline, including a SAM 3.1 step that tightens boxes, and the 36 dataset scores are averaged equally.

MiniMax-M3.1-Flash scored 0.317 mAP50, fifth among the eight paid models costing under $5 per 1,000 images. Claude Haiku 5.5 (0.429), Gemini 3.5 Flash-Lite (0.380), GPT-6 Luna (0.370) and GLM-5.3-Flash (0.341) scored higher. It finished ahead of MiMo-V2.6-Flash (0.277), DeepSeek Flash (0.264) and MiMo-V2.6-Pro (0.239). The hosted Qwen3.8 27B (0.351) and Gemma 4 26B (0.344), which run on Ultralytics' own GPUs without a provider key, also scored higher.

Its recall of 0.413 exceeded its precision of 0.379. Against Gemini 3.5 Flash-Lite, which costs about the same ($1.77 per 1,000 images), it led in aerial and remote sensing (0.32 vs. 0.27), tied on medical images and trailed in nine other domains. Both scored zero on security X-rays. Its best datasets included SSDD satellite radar ship detection (0.86), EgoHands (0.90) and GRAZPEDWRI-DX pediatric wrist X-rays (0.25).

Cost, speed and what it means for users

Most of the $1.84 per 1,000 images comes from reasoning output, since the token price is among the lowest on the Platform but the model thinks before answering. Ultralytics notes that it is not the cheapest route to this accuracy: Claude Haiku 5.5 scored about 0.11 mAP50 higher at about $0.37 per 1,000 images, and GPT-6 Luna (about $0.77) and GLM-5.3-Flash (about $0.99) also scored higher for less.

It is also slower than its price suggests. The median request took about 16 seconds per image, against 1.6 seconds for Gemini 3.5 Flash-Lite and 2.7 seconds for Claude Haiku 5.5. About 1.4% of requests did not finish within the 115-second limit. Such an image comes back with no boxes and can be predicted again.

For users, the addition gives another option, particularly for aerial and remote sensing data. Ultralytics advises treating the boxes as a first draft for human review and checking a few predictions on your own images before annotating a full dataset.