@moondreamai profile picture
moondream @moondreamai

Production vision AI that runs everywhere. Open models, fine-tuning, cloud, and fast inference for real-world deployment.

Joined March 2024
  • Posts 230
  • Following 3
  • Followers 12,206
98 Photos and videos
Pinned Post
moondream
 @moondreamai
Jun 8
Photon 1.3 is out! It's now free and runs Moondream up to 70% faster. More images/sec, lower latency, on Mac, Windows & NVIDIA. Good, fast, cheap: we picked all three! moondream.ai/blog/photon-1-3…

Photon is now free | Moondream

Photon 1.3.0 makes Moondream faster across NVIDIA, Mac, and Windows, runs finetunes on far more hardware, fixes an accuracy issue on older GPUs — and running Moondream locally is now completely free.

moondream.ai
4
3
76
16,922
Time to touch grass and use Moondream. Segmenting a dog is easy. But understanding which dog you mean is a different problem. Moondream does it from a prompt.
0:07
2
6
56
3,165
moondream reposted
Ethan Reid
 @EthanReidMorro
Jun 3
recently had the pleasure of presenting at the Frontier Research Club, hosted by @Stanford. Check out the recording for the TLDR on how we achieved SOTA segmentation
4
2
24
1,967
Learn more about Lens fine-tuning here : moondream.ai/p/lens

Moondream

A fast & powerful vision model that rocks.

moondream.ai
1
1
875
Want to build an agent that can identify the lead swimmer in an Enhanced Games race clip? Existing models don't handle this well. A quick Moondream fine-tune solved it, segmenting the lead swimmer throughout the video.
0:17
3
3
53
7,121
Learn more here: moondream.ai/skills/segment

Segmentation

State-of-the-art open-vocabulary segmentation. Describe any object in natural language and get pixel-accurate SVG masks instantly. No predefined classes, no retraining required.

moondream.ai
5
779
"Succulent chinese meal" isn't an object category. Yet Moondream knows exactly what you mean. Prompt: "croissant", "succulent chinese meal", "pizza", "beer", "check" Pixel-accurate grounding from naturel language. Fast, open vision AI.
0:14
6
6
86
9,656
$5 free credits / month on Moondream spend it on :
2
3
55
3,454
moondream reposted
Ethan Reid
 @EthanReidMorro
May 26
mini tutorial on how to train @moondreamai to watch an NBA game and spot the player with the ball.
11:16
6
5
77
11,483
Even more details: moondream.ai/blog/photon-1-2…

Photon 1.2.0: Faster Inference, Now on Mac, Windows, Blackwell, and Jetson Thor | Moondream

Photon 1.2.0 brings native inference to Apple Silicon and Windows, adds NVIDIA Blackwell and Jetson Thor support, and ships meaningful speed gains across existing GPUs.

moondream.ai
1
7
1,523
Bottom line: Photon 1.2.0 makes Moondream faster, easier to deploy, and available on more of the hardware you use. Production vision AI is moving from “cloud-only” to “run it where the work happens.”
1
6
1,645
Jetson Thor support brings Moondream to robotics, inspection, kiosks, vehicles, cameras, and embedded vision products. Local VLM inference at the edge. More inference timings:
1
7
754
On NVIDIA B200 (server), Photon reaches flagship performance numbers: ~23ms single-request latency for Moondream 2 ~30ms for Moondream 3 Visual reasoning at 30fps speeds.
1
17
1,377
Mac: runs on M-series Macs with native Metal kernels, making local demos, prototypes, desktop apps, and privacy-sensitive workflows much more practical. Windows: No WSL workaround. No Linux wrapper. Runs CUDA inference directly on Windows x86_64. 👇
1
15
1,457
Photon 1.2.0 is here 🚀 ✅ Apple Silicon support ✅ Native Windows CUDA inference ✅ NVIDIA Blackwell support ✅ Jetson Thor edge inference ✅ Faster existing NVIDIA GPU performance Learn more👇
5
11
164
20,948
Base Moondream: 18 percent accuracy on glaucoma detection. GPT-5.4: 33 percent. Moondream plus Lens: 69 percent. Got a Vision AI challenge? Let us show you what Lens can do.
0:41
7
4
126
8,156
Powering the only agentic vision demo at NAB! Moondream counts people over time, captions the scene, and detects the thumbs up people are giving @ptzoptics.
1
26
2,033
Launching Lens, our fine-tuning service for Moondream. Improve Moondream's accuray to production-ready levels in hours, with as few as 20 images. Example: we taught Moondream to find the player with the ball in NBA footage. F1 jumped from 28% to 79%, beating ChatGPT. 54 minutes, $16.89. Fine-tune your own: moondream.ai/blog/lens-moond…
0:41
8
21
317
30,370
moondream reposted
Bruno Pessoa
 @BrunoPessoa22
Apr 18
Cameras that follow the ball are over. PTZOptics Moondream demo at NAB 2026 today: AI that understands who has the ball, when a shot happens, what the play is. Not tracking. Context. Every sports broadcast truck runs a vision-language model in 5 years.
2
8
1,148
Technical whitepaper on Moondream segmentation is out: arxiv.org/abs/2604.02593

Moondream Segmentation: From Words to Masks

We present Moondream Segmentation, a referring image segmentation extension of Moondream 3, a vision-language model. Given an image and a referring expression, the model autoregressively decodes a...

arxiv.org
moondream
 @moondreamai
Mar 11
Moondream segmentation just leveled up. ↑ Better masks ↑ New SOTA benchmarks ↑ 40% faster Already live on Moondream Cloud. Local model technical whitepaper coming later this week. Read more: moondream.ai/blog/segmenting…
1
3
79
9,424
Load more