> they provide a better or at least as good an answer as LLMs in 62.5% of the cases.
Are we going to scrap hospitals because a vet could do the job 62.5% of the time?
The economics also point away from everyone buying a big RAM Mac that sits idle 99% of the time. SLM and own hardware sounds efficient and “free” but it is nothing of the sort when you factor everything in (and forfeit the sharing efficiencies of API)
SLMs are great esp for task specific fine tunes but this take isn’t it
This logic seems mad. If people only need SLMs then hyperscalers can also centrally host higher-efficiency models, and still gain efficiencies of scale and convenience over hosting locally.
The valuations of the hyperscalars won't sustain just being more efficient than something you can run locally. There's a market there, but it's for margin on a commodity. They're priced for oligopoly on unique, premium products.
You have to remember that articles like these are written for finance people who don't understand the underlying technology, by finance people who don't understand the underlying technology. In this case, the author is a "CFA Institute Enterprising Investor", previously a CIO and basically their entire life been "money, money & money", so hardly surprising they're pulling a lot of assumptions based on what they read.
Read the paper the author talks about yourself instead (https://arxiv.org/abs/2511.07885), and also, contrary to what the author says in the article, do not do investments based on single papers made from academic studies, regardless of how much money this guy tells you you can make.
From what’s presented this seems to be the lower end of Q&A and reasoning tasks and not long horizon agentic work. I agree that the search engine replacement AI usage is something that can run anywhere (though it’s still better run in the cloud for speed, context length, sandboxing and convenience) but this isn’t the engine of AI growth.
Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights? They’ll likely be shipped as an ASIC (or MSIC) at that point anyways. Those will use a licensed model from the current leaders. The whole argument sounds like saying that cloud services shouldn’t be profitable because everyone has a computer at home or to meme “we have AI at home”.
This is up to October 2025 though, yes? Obviously things are continually moving but Opus 4.5 launched in November and that was a recognised step change in capability. An up to date comparison would be interesting.
The paper focuses on "intelligence per watt (IPW)", as a way to compare SLMs vs LLMs.
What might happen is that a chunk of the market, whatever its size will be, will end up going to SLMs run on iphones or Macbooks, and eat some of the revenues from LLMs, because not everyone needs the most powerful LLM all the time.
If you are like Google or Apple and you are delivering AI to a mass market unwilling to pay a lot for it, you are absolutely going to drive AI processing to endpoint devices. You are also going to spend what it takes in R&D make a hybrid system that knows when to use local compute or cloud compute. That's going to be the bulk of the workload.
If this is correct I see a future where the hyperscalers are funded by the businesses integrating siloed SLMs in their software.
Also the defence/intelligence industry will always want to keep an edge so don't be surprised if they stick around and we see favourable regulations for them similarly to how the government turns a blind eye to social media platforms because they increase the footprint of mass surveillance.
I wouldn't be surprised if the hyperscalers became software auditors and any piece of critical software was required to have a regulated security audit before it could enter production. Selling the poison and the cure is a great business model.
How many developers here don't see a difference between the latest LLMs and SLMs they can run on their own computer? I tried running a smaller model locally, and it's not usable for me.
I know people like to "predict" things, so that if they happen they can then say "I am a visionary, I predicted it" and start their blog posts with "as I predicted long ago (because I am a visionary), ...".
> The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
If something better comes, it will be better. Sure. And we would like to have something better, because it would be better.
> I tried running a smaller model locally, and it's not usable for me.
If you have the hardware, a MacBook Pro for Qwen 3.6 35B A3B and Gemma 4 26B A4B for example, they are absolutely usable, both in terms of speed and quality. Anecdotally, I can use Qwen for day-to-day coding tasks in TS and Go, without hickups.
I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex, and I have 96GB of VRAM available and tried every local model under the sun so far. Neither of those you mention I'd say are good enough for day to day software engineering for me, but I'm also really strict about code quality and iterate on what outputs agents give me a lot before I'm happy.
With local models, this iteration cycle takes maybe 30 minutes for a single fix or feature, rather than 10 minutes with GPT+Codex, as there is so many corrections and iterations needed, although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
Most of the 27-35B models are LLMs; Gemma 4 12B is an LLM.
The defining characteristic is not really size, IMO — it is the breadth and intent of the training. Is the training set broad or narrow and curated? Is the intent to produce a generalist or a narrow domain specialist?
Size is not a meaningful distinction (on its own at least). Gemma 4 E4B is an edge model but it has the generalist training intent of a large language model.
>> I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
The lack of logic and risk management on this statement, is so strong, I hope humans are all quickly substituted by LLMs. Lets just do it and be done with it...
A remaining advantage of large language models is that as they get larger, they tend to hallucinate less, simply because the odds of the training set containing a desired answer improve with size. If a solid "I don't know" detector is developed for inference, then you can try a small language model first.
An implication is that successful research in "I don't know" detection could destroy hundreds of billions in shareholder value.
Another "cool but we don't know how yet" thing would be a "confidence interval" so we know how much to trust LLM responses. Or while we're fantasizing, they could just know everything all the time regardless of training data. The "if a solid" part is easy to imagine, hard to implement :)
> If their results are true, then we will hardly need any data centres in the future, and the hyperscalers are wasting hundreds of billions of dollars in investments.
Haha, no. They get sufficiently powered and watered industrial warehouses close to where the successful people live. That’s a jackpot for developers, although it destroys the neighborhood as part of the deal.
Cool, they scored well on all the "make complex calculations and I'll vibe check your results based on my own domain experience" things I use the average LLM chatbot for.
So maybe in 10yr I'll be able to run a SLM on a 5yo laptop and not have Google or whoever hoover up everything.
As much as I want local and open-weights models to succeed, nothing beats a paid frontier model for now. Anybody who claims otherwise is simply not a daily user of such models. So this "investor" here should invest sime time in actually using the various LLM models and get a real taste of what it's like.
Their point isn’t that local models are better or even as good more but that if you can do 50%+ of tasks with local then that’s 50% of tokens that aren’t captured as compute done in data centers.
> As you can see, on average, SLMs are as good if not better than LLMs in 81.2% of the cases, with the LLMs having a significant advantage only in areas like engineering, life sciences, transportation and computer sciences.
So what I’m reading here is “LLMs have a significant advantage” in the most critical areas that have practically infinite demand for more intelligence.
The article is kinda dumb, and yes this is clearly the area where frontier models having and advantage matters the most, but I'd point out that these smaller open-weight models are performing better than the big Frontier models of just 4-6 months ago.
This means that the Frontier labs are under immense pressure to maintain that lead, and could end up in serious trouble if they stumble at all.
The other thing id point out is that a lot of us who are token-sensitive do things like build plans using expensive, smart models, and then execute those plans using cheaper dumber models.
While that's true, the open / local models are getting good enough. Given time and the technology trend people may prefer a private local model for most use cases. Nobody is arguing that a Ferrari isn't a faster car, but the Honda is the more practical choice.
You completely missed the thesis here, and that is supported by the numbers being presented. It is that a large share of ordinary inference can be routed away from the hyperscalers.
It is not the whole story, and knowledge is very lacking, but it has gotten a lot of attention. That model together with DeepSeek V4 Flash are the highlights of this summer on the open/local models side.
Complete nonsense.
> they provide a better or at least as good an answer as LLMs in 62.5% of the cases.
Are we going to scrap hospitals because a vet could do the job 62.5% of the time?
The economics also point away from everyone buying a big RAM Mac that sits idle 99% of the time. SLM and own hardware sounds efficient and “free” but it is nothing of the sort when you factor everything in (and forfeit the sharing efficiencies of API)
SLMs are great esp for task specific fine tunes but this take isn’t it
This logic seems mad. If people only need SLMs then hyperscalers can also centrally host higher-efficiency models, and still gain efficiencies of scale and convenience over hosting locally.
The valuations of the hyperscalars won't sustain just being more efficient than something you can run locally. There's a market there, but it's for margin on a commodity. They're priced for oligopoly on unique, premium products.
You have to remember that articles like these are written for finance people who don't understand the underlying technology, by finance people who don't understand the underlying technology. In this case, the author is a "CFA Institute Enterprising Investor", previously a CIO and basically their entire life been "money, money & money", so hardly surprising they're pulling a lot of assumptions based on what they read.
Read the paper the author talks about yourself instead (https://arxiv.org/abs/2511.07885), and also, contrary to what the author says in the article, do not do investments based on single papers made from academic studies, regardless of how much money this guy tells you you can make.
The logic seems mad to me because SLMs can simply not hold as much information as an LLM.
Maybe if you combine an SLM with a database (as a tool) then it could work, but someone should first prove that.
But do you really need a model that has the complete Duran Duran discography memorized and preloaded in RAM at all time?
From what’s presented this seems to be the lower end of Q&A and reasoning tasks and not long horizon agentic work. I agree that the search engine replacement AI usage is something that can run anywhere (though it’s still better run in the cloud for speed, context length, sandboxing and convenience) but this isn’t the engine of AI growth.
Also, the average consumer is not going to be running a local model until they are built into the hardware they already buy and when they are, who is supplying the weights? They’ll likely be shipped as an ASIC (or MSIC) at that point anyways. Those will use a licensed model from the current leaders. The whole argument sounds like saying that cloud services shouldn’t be profitable because everyone has a computer at home or to meme “we have AI at home”.
This is up to October 2025 though, yes? Obviously things are continually moving but Opus 4.5 launched in November and that was a recognised step change in capability. An up to date comparison would be interesting.
The paper focuses on "intelligence per watt (IPW)", as a way to compare SLMs vs LLMs.
What might happen is that a chunk of the market, whatever its size will be, will end up going to SLMs run on iphones or Macbooks, and eat some of the revenues from LLMs, because not everyone needs the most powerful LLM all the time.
If you are like Google or Apple and you are delivering AI to a mass market unwilling to pay a lot for it, you are absolutely going to drive AI processing to endpoint devices. You are also going to spend what it takes in R&D make a hybrid system that knows when to use local compute or cloud compute. That's going to be the bulk of the workload.
Haven't the SLMs been distilled using the LLMs?
If this is correct I see a future where the hyperscalers are funded by the businesses integrating siloed SLMs in their software.
Also the defence/intelligence industry will always want to keep an edge so don't be surprised if they stick around and we see favourable regulations for them similarly to how the government turns a blind eye to social media platforms because they increase the footprint of mass surveillance.
I wouldn't be surprised if the hyperscalers became software auditors and any piece of critical software was required to have a regulated security audit before it could enter production. Selling the poison and the cure is a great business model.
"If", sure.
How many developers here don't see a difference between the latest LLMs and SLMs they can run on their own computer? I tried running a smaller model locally, and it's not usable for me.
I know people like to "predict" things, so that if they happen they can then say "I am a visionary, I predicted it" and start their blog posts with "as I predicted long ago (because I am a visionary), ...".
> The research report estimates that the addressable market in the US for SLMs has grown to about $10tn or one-third of the entire US GDP of $30tn. There isn’t much left for LLMs to thrive in, and every year, their advantage over SLMs is shrinking.
I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
If something better comes, it will be better. Sure. And we would like to have something better, because it would be better.
> I tried running a smaller model locally, and it's not usable for me.
If you have the hardware, a MacBook Pro for Qwen 3.6 35B A3B and Gemma 4 26B A4B for example, they are absolutely usable, both in terms of speed and quality. Anecdotally, I can use Qwen for day-to-day coding tasks in TS and Go, without hickups.
I'm unable to find a local model that comes close to the effectiveness of GPT models in Codex, and I have 96GB of VRAM available and tried every local model under the sun so far. Neither of those you mention I'd say are good enough for day to day software engineering for me, but I'm also really strict about code quality and iterate on what outputs agents give me a lot before I'm happy.
With local models, this iteration cycle takes maybe 30 minutes for a single fix or feature, rather than 10 minutes with GPT+Codex, as there is so many corrections and iterations needed, although I will say that the speed I'm able to get locally makes it more fun that any of the remote models.
You let a hiccup slip through in your comment though.
Most of the 27-35B models are LLMs; Gemma 4 12B is an LLM.
The defining characteristic is not really size, IMO — it is the breadth and intent of the training. Is the training set broad or narrow and curated? Is the intent to produce a generalist or a narrow domain specialist?
Size is not a meaningful distinction (on its own at least). Gemma 4 E4B is an edge model but it has the generalist training intent of a large language model.
>> I stopped counting the number of times "estimates" said that a market would absolutely explode, and it absolutely didn't. Those are in the business of being a broken clock.
The lack of logic and risk management on this statement, is so strong, I hope humans are all quickly substituted by LLMs. Lets just do it and be done with it...
> Lets just do it and be done with it...
Presumably not what you intended but this phrase immediately takes me to:
https://www.youtube.com/watch?v=dJFR7xbOIuw&t=42s
Great movie...yeah I think I was inspired by the scene... :-)
> I tried running a smaller model locally, and it's not usable for me.
Probably a skill issue on your part.
A remaining advantage of large language models is that as they get larger, they tend to hallucinate less, simply because the odds of the training set containing a desired answer improve with size. If a solid "I don't know" detector is developed for inference, then you can try a small language model first.
An implication is that successful research in "I don't know" detection could destroy hundreds of billions in shareholder value.
>> A remaining advantage of large language models is that as they get larger, they tend to hallucinate less
First time I hear that...not really true.
"Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors" - https://arxiv.org/abs/2607.00447
"Calibrated Language Models Must Hallucinate" - https://arxiv.org/abs/2311.14648
"TruthfulQA: Measuring How Models Mimic Human Falsehoods" - https://arxiv.org/abs/2109.07958
Another "cool but we don't know how yet" thing would be a "confidence interval" so we know how much to trust LLM responses. Or while we're fantasizing, they could just know everything all the time regardless of training data. The "if a solid" part is easy to imagine, hard to implement :)
> If their results are true, then we will hardly need any data centres in the future, and the hyperscalers are wasting hundreds of billions of dollars in investments.
Haha, no. They get sufficiently powered and watered industrial warehouses close to where the successful people live. That’s a jackpot for developers, although it destroys the neighborhood as part of the deal.
Cool, they scored well on all the "make complex calculations and I'll vibe check your results based on my own domain experience" things I use the average LLM chatbot for.
So maybe in 10yr I'll be able to run a SLM on a 5yo laptop and not have Google or whoever hoover up everything.
As much as I want local and open-weights models to succeed, nothing beats a paid frontier model for now. Anybody who claims otherwise is simply not a daily user of such models. So this "investor" here should invest sime time in actually using the various LLM models and get a real taste of what it's like.
Their point isn’t that local models are better or even as good more but that if you can do 50%+ of tasks with local then that’s 50% of tokens that aren’t captured as compute done in data centers.
> As you can see, on average, SLMs are as good if not better than LLMs in 81.2% of the cases, with the LLMs having a significant advantage only in areas like engineering, life sciences, transportation and computer sciences.
So what I’m reading here is “LLMs have a significant advantage” in the most critical areas that have practically infinite demand for more intelligence.
The article is kinda dumb, and yes this is clearly the area where frontier models having and advantage matters the most, but I'd point out that these smaller open-weight models are performing better than the big Frontier models of just 4-6 months ago.
This means that the Frontier labs are under immense pressure to maintain that lead, and could end up in serious trouble if they stumble at all.
The other thing id point out is that a lot of us who are token-sensitive do things like build plans using expensive, smart models, and then execute those plans using cheaper dumber models.
While that's true, the open / local models are getting good enough. Given time and the technology trend people may prefer a private local model for most use cases. Nobody is arguing that a Ferrari isn't a faster car, but the Honda is the more practical choice.
You completely missed the thesis here, and that is supported by the numbers being presented. It is that a large share of ordinary inference can be routed away from the hyperscalers.
How does a current local model compare to the best frontier model 12 months ago. Or 24 months ago?
It beats a frontier model from 12 months according to this bench: https://news.ycombinator.com/item?id=49334544
It is not the whole story, and knowledge is very lacking, but it has gotten a lot of attention. That model together with DeepSeek V4 Flash are the highlights of this summer on the open/local models side.
I have found qwen 3.8's coding quality using opencode to be similar to claude or gpt from 6-9 months ago, except much slower.