
doing whatever it takes to bring local AI to every phone and laptop. Previously @Meta
Manhattan
Joined August 2026
- Tweets375
- Following566
- Followers1.1K
- Likes2K
Pinned Tweet
Local AI caught up way faster than anyone realizes. Sharing my thoughts having spent time in ML at Meta and dedicating the last year to getting smart on local AI.
1/ Barely anyone, not even hardcore AI power users, realizes how fast local models have caught up to the frontier. Qwen3.8 woke a few people up, but the most technical people are so frontier-pilled they've blinded themselves to the breakneck progress.
Meanwhile, normal users don't care about frontier intelligence at all. They want AI that gets the job done, but they have no idea what a local model is or how to run one.
So local is stuck in a weird spot: the people who can run these models don't care to do so, and the people who'd benefit most don't know how to run local.
Good on Yann for calling anthropic's BS out.
We're at a sharp fork in the road. One where incredible intelligence becomes a freely accessible commodity that anyone can own, and another highly dystopian path where two mega corporations control all intelligence in perpetuity.
I'll always fight for the former version to play out.
Replying to @huxlab
Dario's spiel since GPT-2 has always been about hyping the dangers so that open models would be made illegal and *he* would stay in control.
He left OpenAI because he wasn't going to have control.
It's like the Ottoman Empire banning the printing press for Arabic so they would keep in control of the dogma.
It stops the population from freely accessing diverse knowledge.
It's a new form of medieval obscurantism.
SF is the most superficial city I’ve ever spent time in.
The first question anyone asks is what you do for work. Then, where you went to school. Depending on the first two answers, they might ask your name.
I’ve seen tweets with a couple hundred likes with 600k imps but this one’s stuck at ~3k. Not cool, Dario 😔
With American labs now taking the lead in advancing open local models, it will become increasingly difficult for Ant/OpenAI to paint open models as an existential threat to Americans (read: to their business).
Couldn't have come at a better time
Thanks @miramurati
1
15
783
Data center AI and laptop/local AI are forking into two different architectures, and it's long overdue.
data center serving infra has lots of fast memory. Laptops, not so much.
the best models will be the ones that adapt to the hardware they are served on.
looping transformers are a great adaptation for a laptop, and soon, for phones.
Inkling 2 Mini is a great example of what's to come - its middle layers run 3x, so with the same weights, it gets much deeper reasoning than a non-looped model that's the same size.
Translation: “we distilled the entire open internet, and we’re now pulling up the ladder behind us”
Distillation can reproduce powerful AI capabilities without the safeguards built around them.
We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts.
These findings are helping us sharpen our defenses and support safeguards across the industry.
openai.com/index/combatting-…
Michael Waldman retweeted
i am so so so so so so so so so so excited about local models, open weights, private inference, etc.
it reminds me of the early days of Bitcoin when the focus was on sovereignty and censorship resistance.
where are the philosophers of this movement? must follow!
Michael Waldman retweeted
Nobody is distilling your models buddy
Distillation can reproduce powerful AI capabilities without the safeguards built around them.
We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts.
These findings are helping us sharpen our defenses and support safeguards across the industry.
openai.com/index/combatting-…
one thing I haven't seen enough people talking about regarding inkling 2 mini which is super encouraging:
On release day, open models tend to run sub-optimally for some time until inference engines develop architecture-specific optimizations that speed them up several multiples. We saw this with Qwen 3.8.
TLDR: expect this model to get much faster over the next few weeks and months. Just the beginning
With American labs now taking the lead in advancing open local models, it will become increasingly difficult for Ant/OpenAI to paint open models as an existential threat to Americans (read: to their business).
Couldn't have come at a better time
Thanks @miramurati
Let's put this into perspective.
When Qwen 3.8 came out, it was a laptop-sized model that beat the Anthropic flagship (opus 4.6) of 6 months ago.
@thinkymachines just released a model that is 5x faster than Qwen3.8 while being 15% more accurate.
It's a 26b that's 12x faster than Ripple's 65b.
People have debated if local would depend on Chinese models, TM just answered that with a definitive no.
Months ago, no one was sure if coding would be tenable locally, even years out. This release answered that question very clearly as well.
Local steps up another order of magnitude. The frontier is moving onto your machine.
Nikita is absolutely right.
We counterintuitively chose to focus first on phones over PCs, even though at the time the models weren’t ready.
We knew that the models that will be good enough on a phone will release over the next year, and have already seen it with models like @liquidai LFM3.
What those models need is an inference engine and harness optimized around them, that can get the most out of a phone, and manage constraints like thermals and lower memory efficiently.
If you’d like to experience owning AI on your phone, check out @askloci - trusted by tens of thousands and growing daily.
There would be a massive market opportunity unlocked if there were good local models that could run on a mobile device.
There are vast amounts of data on our phones that is too private and too large to give to a frontier model—especially in today’s cybersecurity environment.
The benefits may not accrue to the model developer but it would be a big win for society.
Michael Waldman retweeted
The implications of this is basically:
You can now run Opus 5.5 on a high end macbook
In 6 months opus 5.5 will run on a shit laptop
In a year you’ll have fable locally on your phone.
Is paying $200/month for a cloud AI that steals all your info a thing of the past?
Let's put this into perspective.
When Qwen 3.8 came out, it was a laptop-sized model that beat the Anthropic flagship (opus 4.6) of 6 months ago.
@thinkymachines just released a model that is 5x faster than Qwen3.8 while being 15% more accurate.
It's a 26b that's 12x faster than Ripple's 65b.
People have debated if local would depend on Chinese models, TM just answered that with a definitive no.
Months ago, no one was sure if coding would be tenable locally, even years out. This release answered that question very clearly as well.
Local steps up another order of magnitude. The frontier is moving onto your machine.
Let's put this into perspective.
When Qwen 3.8 came out, it was a laptop-sized model that beat the Anthropic flagship (opus 4.6) of 6 months ago.
@thinkymachines just released a model that is 5x faster than Qwen3.8 while being 15% more accurate.
It's a 26b that's 12x faster than Ripple's 65b.
People have debated if local would depend on Chinese models, TM just answered that with a definitive no.
Months ago, no one was sure if coding would be tenable locally, even years out. This release answered that question very clearly as well.
Local steps up another order of magnitude. The frontier is moving onto your machine.
Michael Waldman retweeted
oooh finally a new 26B model! any mac with 32GB+ RAM can run it
i'm quite bullish about the future of these small LLMs - they will continue to get better and eventually able to support a wide range of tasks, and we can all fine-tune them easily for our needs
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Michael Waldman retweeted
BREAKING: Jensen Huang reportedly painted a fake tunnel on a brick wall, tricking an employee into running headfirst into it.
At first glance this is a CRAZY release...
Testing out now, this model may have dethroned Qwen3.8 for best laptop-size model, and it's coming from an American lab too.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Cognition is one of the most talent dense companies.
Why is adding support for one MCP a launch announcement?
✨ Introducing DevinDash ✨
Starting today, dash food with the DoorDash MCP and a Devin Bot.
Startups can claim $1k of free DoorDash: devin.ai/startups
26
79
12,176

















