Everything we announced at our first LlamaCon
57 comments
·April 29, 2025throwaw12
Feels like Meta is going to Cloud services business but in AI domain. They resisted entering cloud business for so long, with the success of AWS/Azure/GCP I think they are realizing they can't keep at the top only with social networks without owning a platform (hardware, cloud)
Keyframe
If Lidl can venture into cloud business, I guess so can Meta.
paxys
In this case the market basically validated itself. Companies are already using Llama for production workloads. It is offered as a first class LLM option in AWS, Azure, GCP and all other major hosting providers. Meta may have been getting marginal licensing fees out of it but now wants a bigger piece of the pie.
throwaw12
SAM 3 (Segment Anything Model) is coming this summer
daemonologist
SAM's a really cool model, that's something to look forward to. I didn't see that in the LlamaCon notes, is that something they've announced elsewhere or just a rumor atm?
ipsum2
It was mentioned briefly. https://ai.meta.com/sam3
swatcoder
They seem to see tge writing on the wall and have been panicked for a while, yes.
Gobbling up rising brands kept their finances going for a while, but the grand Metaverse pivot was clearly their (much struggling) attempt to invent their own titanic platform akin to Android or iPhone.
With that not gaining as much traction as they wanted as quickly as they wanted, they're still on the hunt, as here.
retinaros
The metaverse is a great idea but they should have partnered with Epic for this or Valve. The implementation was subpar
mohsen1
I am guessing because of Qwen 3 release they pulled back the reasoning model that was likely due to launch today.
scosman
Anyone manage to sign up for the waitlist? I just get a redirect loop back to the login when requesting access.
ahmedfromtunis
Does anyone use llama as their primary model for any usecase? Maybe it's my fault for not spending much time with it, but I still couldn't find the applications for which llama has an advantage over the competition.
philipkglass
I recently needed to classify thousands of documents according to some custom criteria. I wanted to use LLM classification from these thousands of documents to train a faster, smaller BERT (well, ModernBERT) classifier to use across millions of documents.
For my task, Llama 3.3 was still the best local model I could run. I tried newer ones (Phi4, Gemma3, Mistral Small) but they produced much worse results. Some larger local models are probably better if you have the hardware for them, but I only have a single 4090 GPU and 128 GB of system RAM.
galeos
How did you find ModernBERT performance Vs prior BERT models?
philipkglass
I didn't try original BERT at all because I didn't get good results from any LLMs on small document excerpts, so I assumed that a substantial context was necessary for good results. Traditional BERT only accepts up to 512 tokens, while ModernBERT goes up to 8192. I ended up using a 2048 token limit.
mvieira38
It's pretty popular in the local LLM space
littlestymaar
It used to, but Llama 4 is useless for local LLM for most people.
wewewedxfgdf
Can someone explain to me please why Meta doesn't create subject specific versions of their LLMs such as one that knows only about computer programming, computers, hardware software.
I would have imagined such a thing would be smaller and thus run on smaller configurations.
But since I am only a layman maybe someone can tell me why this isn't the case?
ntonozzi
One of the weirdest and most interesting parts of LLMs is that they grow more effective the more languages and disciplines they are trained in. It turns out training LLMs on code instead of just prose boosted their intelligence and reasoning capabilities by huge amounts.
gardenhedge
Source? Sounds interesting
cube2222
Generally, all that non-tech content still helps the model “to learn”.
Also, the software you’re working on will generally in some way have a real-world domain - without knowing it the AI all likely be a less effective assistant. Design conversations with it would likely be pretty non-fun, too.
Finally, the “bitter lesson” article[0] from a couple years ago is I think somewhat applicable too.
[0]: http://www.incompleteideas.net/IncIdeas/BitterLesson.html
bbatha
To add on to the sibling. Specialized models, including fine tuned ones, continually have their lunch eaten by general models within 3-6 months. This time round is mixture of experts that’ll do it, next year it’ll be something else. Tuned models are expensive to produce and are benchmark kings but less do less well in the real world qualitative experience. The juice just ain’t worth the squeeze most of the time.
Meta does have some specialized models though, llamaguard was released for llama 2 and 3.
KTibow
Other companies have done this (see Qwen Coder). It doesn't scale past a few disciplines like math and code though, and using mixtures of experts give you most of the same benefits.
hedayet
Facebook did a great job open sourcing Llama and pushing the market to being competitive, but this list seems super shallow.
0. Introducing Llama API in preview
This one is good but not centre stage worthy. Other [closed] models have been offering this for a long time.
1. Fast inference with Llama API
How fast? and how must faster than others? This section talks about latency and there's absolutely no numbers in this section!
2. New Llama Stack integrations
Speculations with 0 new integration. Llama Stack with NVIDIA had already been announced and then this section ends with '...others on new integrations that will be announced soon. Alongside our partners, we envision Llama Stack as the industry standard for enterprises looking to seamlessly deploy production-grade turnkey AI solutions.'
3. New Llama Protections and security for the open source community
This one is not only the best on this page, but is actually good with announcement of - Llama Guard 4, LlamaFirewall, and Llama Prompt Guard 2
4. Meet the Llama Impact Grant recipients
Sorry but neither the gross amount $1.5 million USD, nor the average $150K/recipients is anything significant at Facebook scale.
amusingimpala75
Meta needs to stop open-washing their product. It simply is not open-source. The license for their precompiled binary blob (ie model) should not be considered open-source, and the source code (ie training process / data) isn’t available.
michaelt
> the source code (ie training process / data) isn’t available
The training data is all scraped from the internet, ebooks from libgen, papers from Sci-Hub, and suchlike.
They don't have the right to redistribute it.
observationist
They've painted themselves into a corner - the second people see the announcement that they've enforced the license on someone, people will switch to actual open source licensed models and Meta's reputation will take a hit.
It's ironic that China is acting as a better good faith participant in open source than Meta. I'm sure their stakeholders don't really care right now, but Meta should switch to Apache or MIT. The longer they wait the more invested people will be and the more intense the outrage when things go wrong.
piperswe
Applying Apache or MIT to a binary blob doesn't make it open source either
bbayer
This is actually my first impression while I am reading the post. Mentions "open source" everywhere but dude how the earth it is open source without training data.
ronsor
Almost no company is going to release training data because they don't want to waste time with lawsuits. That's why it doesn't happen. Until governments fix that issue, I don't even think the "it's not really open without training data!!!" argument is worth any time. It's more worth focusing on the various restrictions in the LLaMA license, or even better, questioning whether model weights can be licensed at all.
logicchains
No new model? Maybe after the Qwen 3 release today they decided to hold back on Llama 4 Thinking until it benchmarks more competitively.
smcnally
Beyond solid benchmarks, Alibaba's power move was dropping a bunch of models available to use and run locally today. That's disruptive already and the slew of fine tunes to come will be good for all users and builders.
https://huggingface.co/collections/Qwen/qwen3-67dd247413f0e2...
walterbell
What's the minimum GPU/NPU hardware and memory to run Qwen3 locally?
Havoc
There is a 0.6B model so basically nothing.
And the MoE 30B one has a decent shot at running OK without GPU. I'm on a 5800x3d so two generations old and its still very usable
smcnally
`model.safetensors` for Qwen3-0.6B is a single 1.5GB file.
Qwen3-235B-A22B has 118 `.safetensors` files at 4GB each.
There are a bunch of models and quants between those.
laweijfmvo
I'm running 4B on my 8GB AMD 7600 via ollama
dpe82
Qwen3 is a family of models, the very smallest are only a few GB and will run comfortably on virtually any computer of the last 10 years or recent-ish smart phone. The largest - well, depends how fast you want it to run.
littlestymaar
There are models down to 0.6B and you can even run Qwen3 30B-A3B reasonably fast on CPU only.
paxys
They released the Llama 4 suite three weeks ago.
Havoc
Unlucky timing for meta...
littlestymaar
It's not about luck, pretty sure that Qwen intentionally bullied them.
yapyap
Was there a ball pit
oofbaroomf
Yeah, it was for the Llama team because they love playing in ball pits instead of releasing good models.
retinaros
did I read well that they have a gated 3.3 8b?
Its impressive that Llama and the Ai teams in general survived the meta-verse push at Facebook. Congrats to the team for keeping their heads down and saving the company from itself.
Its all Ai all the time now though, not seen any mention of our reimagined future of floating heads hanging out together in quite some time.