With SAM 3 you can use text and visual prompts to precisely identify, segment, and follow any object in images or videos—coming soon to Instagram Edits and Vibes on the Meta AI app.
Read the original article
Comments
I’m thankful that Meta still contributes to open source and shares models like this. I know there’s several reasons to not like the company, but actions like this are much appreciated and benefit everyone.
Does everyone forget 2023 when someone leaked the llama weights to 4chan?? Then meta started issuing takedowns on the leaks trying to stop it.
Meta took the open path because their initial foray into AI was compromised so they have been doing their best to kneecap everyone else since then.
I like the result but let’s not pretend it’s for gracious intent.
By prodigycorp 2025-11-2013:451 reply Wait a minute. I'm no Meta fan, but that leak wasn't internal. llama released their weights to researchers first. The leak was from the initial batch of users, not from inside of Meta. iirc, the model was never meant to be closed weight.
By poutrathor 2025-11-2019:50 I agree How can the previous comment be on hacker news ? Every one here has followed the llama release saga. The famous cheeky PR on their GitHub with the torrent link was genius comedy.
By zamadatix 2025-11-2013:16 This might make sense for explaining n=1 releases of Llama being open weight. Even OpenAI started with open weight models and moved to closed weight though, so why would this have forever locked Meta into releasing all models as open weight and across so many model families if they weren't really interested in that path as a strategy in its own right?
There is so much malice in the world, let’s just pretend for once it is gracious intent. Feels better.
Sure, but people have the right to ask questions, as for example Zuck's pledge to give away 99% which people pointed out might be a tax avoidance scheme
The retort was essentially "Can't you just be nice?" but people have the right to ask questions; sometimes the questions reveal much corruption that actually does go on
By visioninmyblood 2025-11-205:073 reply Not of fan of the company for the social media but have to appreciate all the open sourcing. none of the other top labs release thier models like meta.
By magicalist 2025-11-2017:50 > none of the other top labs release thier models like meta
Don't basically all the "top labs" except Anthropic now have open weight models? And Zuckerberg said they were now going to be "careful about what we choose to open source" in the future, which is a shift from their previous rhetoric about "Open Source AI is the Path Forward".
They're not doing it out of the goodness of their heart, they're deploying a classic strategy known as "Commoditize Your Complement"[1], to ward off threats from OpenAI and Anthropic. It's only a happy accident that the little guy benefits in this instance.
Facebook is a deeply scummy company[2] and their stranglehold on online advertising spend (along with Google) allows them to pour enormous funds into side bets like this.
By unsungNovelty 2025-11-2011:02 Not even closely OK with facebook. But none of the other companies do this. And Mark has been open about it. I remember him saying in an interview the same very openly. Something oddly respectable about NOT sugar coating with good PR and marketing. Unlike OpenAI.
By arcanemachiner 2025-11-209:351 reply Well, when your incentives happen to align with those of a faceless mega-corporstion, you gotta take what you can get.
By jayd16 2025-11-2016:08 We can still like it. We're not nominating Nobel Prizes or something.
Among the top 10 tech companies and beyond, they have the most successful open source program.
These projects come to my mind:
SAM segment anything.
PyTorch
LLama
...
Open source datacenters and server blueprints.
the following instead comes from grok.com
Meta’s open-source hall of fame (Nov 2025)
---------------------
Llama family (2 → 3.3) – 2023-2025 >500k total stars · powers ~80% of models on Hugging Face Single-handedly killed the closed frontier model monopoly
---------------------
PyTorch – 2017 85k+ stars · the #1 ML framework in research TensorFlow is basically dead in academia now
---------------------
React + React Native – 2013/2015 230k + 120k stars Still the de-facto UI standard for web & mobile
---------------------
FAISS – 2017 32k stars · used literally everywhere (even inside OpenAI) The vector similarity search library
---------------------
Segment Anything (SAM 1 & 2) – 2023-2024 55k stars Revolutionized image segmentation overnight
---------------------
Open Compute Project – 2011 Entire open-source datacenter designs (servers, racks, networking, power) Google, Microsoft, Apple, and basically the whole hyperscaler industry build on OCP blueprints
---------------------
Zstandard (zstd) – 2016 Faster than gzip · now in Linux kernel, NVIDIA drivers, Cloudflare, etc. The new compression king
---------------------
Buck2 – 2023 Rust build system, 3-5× faster than Buck1 Handles Meta’s insane monorepo without dying
---------------------
Prophet – 2017 · 20k stars Go-to time-series forecasting library for business
---------------------
Hydra – 2020 · 9k stars Config management that saved the sanity of ML researchers
---------------------
Docusaurus – 2017 · 55k stars Powers docs for React, Jest, Babel, etc.
---------------------
Velox – 2022 C++ query engine · backbone of next-gen Presto/Trino
---------------------
Sapling – 2023 Git replacement that actually works at 10M+ file scale
---------------------
Meta’s GitHub org is now >3 million stars total — more than Google + Microsoft + Amazon combined.
---------------------
Bottom line: if you’re using modern AI in 2025, there’s a ~90% chance you’re running on something Meta open-sourced for free.
By unsungNovelty 2025-11-2010:58 I dont think it's open source. It says SAM license. Most likely source available.
By uudecoded 2025-11-206:51 Agreed. The community orientation is great now. I had mixed feelings about them after finding and reporting a live vuln (medium-severity) back in 2005 or so.[1] I'm not really into social media but it does seem like they've changed their culture for the better.
[1] I didn't take them up on the offer to interview in the wake of that and so it will be forever known as "I've made a huge mistake."
By siva7 2025-11-2014:40 If they really deliver a model that can track and describe existing images / videos well that would be a huge breakthrough. There are many extremely useful cases in med, law, surveillance, software and so on. Their competition sucks at this.
By throwuxiytayq 2025-11-2013:24 Disappointingly, every time Zuck hands out some free shit people instantly forget that he and his companies are a cancer upon humanity. Come on dude, "several reasons to not like the company" doesn't fucking cut it.
By daemonologist 2025-11-1921:332 reply First impressions are that this model is extremely good - the "zero-shot" text prompted detection is a huge step ahead of what we've seen before (both compared to older zero-shot detection models and to recent general purpose VLMs like Gemini and Qwen). With human supervision I think it's even at the point of being a useful teacher model.
I put together a YOLO tune for climbing hold detection a while back (trained on 10k labels) and this is 90% as good out of the box - just misses some foot chips and low contrast wood holds, and can't handle as many instances. It would've saved me a huge amount of manual annotation though.
As someone that works on a platform users have used for labeling 1B images, I'm bullish SAM 3 can automate at least 90% of the work. Data prep is flipped to models being human-assisted instead of humans being model-assisted (see "autolabel" https://blog.roboflow.com/sam3/). I'm optimistic majority of users can now start deploying a model to then curate data instead of the inverse.
By darig 2025-11-204:54 [dead]
By pierrec 2025-11-205:49 I'm guessing you worked on the Stokt app or something similar! It's certainly become one of the best established apps in climbing.
The 3D mesh generator is really cool too: https://ai.meta.com/sam3d/ It's not perfect, but it seems to handle occlusion very well (e.g. a person in a chair can be separated into a person mesh and a chair mesh) and it's very fast.
It's very impressive. Do they let you export a 3D mesh, though? I was only able to export a video. Do you have to buy tokens or something to export?
I couldn't download it. Model appears to be comparable to Sparc3D, Huyunan, etc but w/o download, who can say? It is much faster though.
By visioninmyblood 2025-11-1922:47 you can download it at https://github.com/facebookresearch/sam3. for 3d https://github.com/facebookresearch/sam-3d-objects
I actually found the easiest way was to run it for free to see if it works for my use case of person deidentification https://chat.vlm.run/chat/63953adb-a89a-4c85-ae8f-2d501d30a4...
By WhiteNoiz3 2025-11-1923:101 reply The models it creates are gaussian splats, so if you are looking for traditional meshes you'd need a tool that can create meshes from splats.
By modeless 2025-11-1922:11 The model is open weights, so you can run it yourself.
