What is Moondream 3.1 (9B-A2B)?
Moondream 3.1 is M87 Labs' current vision-language model, announced on 7 July 2026. It is a mixture-of-experts model with 9 billion parameters in total and about 2 billion active per token, and it replaces the older Moondream 2 and the Moondream 3 preview as the company's flagship. It reads images and writes text.
It has four built-in skills, each returning structured output: query (answer a question about an image), caption, detect (bounding boxes for any object you name) and point (coordinates). M87 trained a separate LoRA adapter for each benchmark with fine-tuning and reinforcement learning, then distilled them back into the base weights.
The weights are under the Moondream Model License 1.0, not an open-source licence. It allows commercial use, self-hosting, fine-tuning and redistribution with no revenue cap, but running a general-purpose hosted Moondream API for others needs a separate commercial licence. The older Moondream 2 still draws about 2.3 million Hugging Face downloads a month; 3.1 drew about 9,600.
Is Moondream 3.1 (9B-A2B) worth it?
Moondream 3.1 wins on throughput first and accuracy second. In M87's own test on one H100 at batch size 16, it served 34.2 requests a second against 1.44 for Qwen3.5 9B, 4.78 for SAM 3 and 0.46 for Gemma 4 12B, while averaging 77.9 across its benchmark set to their 74, 62.9 and 63. That is roughly seven times the next-fastest model.
Detection is where it leads. It scores 81.46 on COCO against 75.58 for Qwen3.5 9B and 77.06 for SAM 3, 74.61 on the dense Dense200 set against 48.23, 52.77 on crowded retail shelves (SKU-110K) against 44.27, and 59.63 on aerial DOTA-v2 against 50.36. It also reads charts well, at 86.01 on ChartQA against 70.72 for Qwen.
It does not win everything. Qwen3.5 9B beats it on LVIS (71.11 to 67.4), HumanRef, CountBench and DocVQA (92.94 to 88.6), and SAM 3 is better on CrowdHuman. All the numbers come from M87's release post, run with its own Photon engine, and we found no independent reproduction. For reading documents, a dedicated OCR model is still the safer pick.
The licence is the other thing to read before building. Shipping Moondream inside a product, including a SaaS app, is allowed. Reselling open-ended Moondream inference, captioning or visual Q&A as an API is not, even for free, without a commercial deal.
Where it earns its place
A computer-vision engineer at a grocery retail chain
Count products and gaps on shelf photos from every store each night
Call detect with each product name through the Batch API at half the $0.30 and $1.00 per-million rates, or run Photon on one GPU. Its 52.77 on SKU-110K against 44.27 for Qwen3.5 9B is the reason to choose it.
A QA lead at an electronics manufacturer
Flag missing components on circuit-board images on the factory line
Fine-tune a LoRA and run it locally under the Moondream licence, which allows domain-specific products. Point at about 145 ms end to end on Cloudflare, per M87, is fast enough for line-side checks.
A GIS analyst at a regional planning office
Count vehicles and buildings in aerial imagery
Use detect on tiles; it scores 59.63 on DOTA-v2 against 50.36 for Qwen3.5 9B in M87's table. Check counts on a sample, since Qwen still leads CountBench.
Before you commit
- Setup
- pip install moondream, then md.vl(local=True) downloads the weights and runs them with Photon on NVIDIA Ampere or newer GPUs or Apple Silicon. No API key is needed for the base model locally.
- Integrations
- Python client with caption, query, detect and point; Moondream Cloud and Cloudflare Workers AI (@cf/moondream/moondream3.1-9B-A2B) share the same API; Triton for production serving; Lens API for fine-tuning.
- Data and privacy
- Self-hosting keeps images on your hardware. The Moondream Model License allows commercial use and fine-tuning, requires licence notices on redistribution, and bans general-purpose hosted inference without a commercial licence.
- Cost at scale
- Hosted: $0.30 per million input tokens and $1.00 per million output tokens, halved through the Batch API. Self-hosted: one H100 handled 34.2 requests a second in M87's benchmark.
- Limits
- Text output only, no image generation or video. Benchmarks are self-reported. The licence ends automatically on breach unless fixed within 30 days.
What’s good
- 34.2 requests a second on one H100, about 7 times the next model
- Leads COCO, Dense200, SKU-110K and DOTA-v2 detection (self-reported)
- Detect and point skills return structured boxes and coordinates
- Cloud output price cut to $1.00 per million tokens
What’s not
- Custom licence bars general-purpose hosted APIs
- Trails Qwen3.5 9B on DocVQA, LVIS and CountBench
- Benchmarks are M87's own, on its own engine
- Context window not published
Best for
Engineering teams that need open-vocabulary detection, pointing or counting on many images a second, on their own GPUs or a cheap API.
Moondream 3.1 (9B-A2B) alternatives
Moondream 3.1 (9B-A2B): common questions
- Is Moondream 3.1 (9B-A2B) worth it?
- For finding and counting things in images at volume, Moondream 3.1 is faster and more accurate than the general vision models people usually reach for. We rate it 4.2 out of 5.
- How much does Moondream 3.1 (9B-A2B) cost?
- Free weights under the Moondream Model License. Moondream Cloud and Cloudflare Workers AI charge $0.30 per million input tokens and $1.00 per million output tokens, with 50% off through the Batch API.
- What are the downsides of Moondream 3.1 (9B-A2B)?
- Custom licence bars general-purpose hosted APIs; trails Qwen3.5 9B on DocVQA, LVIS and CountBench; benchmarks are M87's own, on its own engine; context window not published.
- Who is Moondream 3.1 (9B-A2B) best for?
- Engineering teams that need open-vocabulary detection, pointing or counting on many images a second, on their own GPUs or a cheap API.
- How much does Moondream 3.1 (9B-A2B) cost for a team or in production?
- Hosted: $0.30 per million input tokens and $1.00 per million output tokens, halved through the Batch API. Self-hosted: one H100 handled 34.2 requests a second in M87's benchmark.
- Is Moondream 3.1 (9B-A2B) safe for company data?
- Self-hosting keeps images on your hardware. The Moondream Model License allows commercial use and fine-tuning, requires licence notices on redistribution, and bans general-purpose hosted inference without a commercial licence.
- What are the best Moondream 3.1 (9B-A2B) alternatives?
- The closest alternatives we have reviewed are SAM 3 (Segment Anything Model 3), LocateAnything-3B, Molmo2-8B and Qwen3.8-27B.
Answers checked Nov 7, 2026.
Make Moondream 3.1 (9B-A2B)? Show this rating on your site
