Rendered at 23:45:06 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
jmward01 3 hours ago [-]
One major consequence of the ramapocalypse, I think, is an even higher focus on small efficient models. I personally believe that the multi-trillion parameter models are fundamentally missing things and the push to smaller, more efficient will drive evolutionary structural changes that will lead to future gains
cogman10 2 hours ago [-]
I'd assume the closed weight models are all working on shrinking their parameter counts anyways. They too benefit from smaller models. It'd be foolish for these SOTA labs to not be working at reducing parameter counts.
cootsnuck 2 hours ago [-]
I would say even without rampocalypse there would still be the strong incentive to innovate at the edge and under more extreme constraints. The incentives are just even stronger now.
I'm looking forward to seeing what types of new things people create over the coming years once there is less obsession with massive unwieldy LLMs. I think the incentives are just too strong to ignore.
XCSme 1 hours ago [-]
This is just temporary though, right?
With the benefit of LLMs already being proven, in a couple of years we will have vastly better hardware for inference I guess.
I feel like now hardware is stagnating a bit, because the software side has moved too fast for the hardware to catch up. Once we settle on some good, optimal software architecture for the models, dedicated hardware will easily increase throughout by 10x or 100x, for a fraction of the cost.
LLMs seems quite simple, maybe we'll be able to print/assemble at home our own chips with the desired models/weights.
Maybe we'll have model weights being shared like game cartridges.
jmward01 49 minutes ago [-]
I personally think of this like sorting algorithms. Quick sort does the same thing bubble sort does so why do we need quick sort? Pushing for efficiency drives innovation. It does this for many reasons but a big one is that putting a cap on a resource forces you to consider the others available and often you find that all it took was a little effort and suddenly the alternate path that looked a little worse is actually better than you realized.
This has a lot to do with how MCTS works BTW. The current best path is often only the current best path because a lot of investment has been sunk into it. If you were to put equal resources into a different path you may find that it was actually far better. It is just that the early rollouts favored the other 'best path' so you sunk a lot of resources into that one. We are very early in our exploration of LLM architecture. I highly doubt we are anywhere near the best path right now.
XCSme 39 minutes ago [-]
Definitely, LLMs are highly ineficient now.
The diffusion models are interesting, but those also seem hacky.
I think the next form of AIs will be simpler and more abstract.
The building blocks of our brain don't have the notion of a "token" embed into them, it's lower level that that.
I think first step is to find a better way to represent information.
LLMs shouldn't "compute" stuff using language tokens, but some other, more efficient logical mechanisms. LLMs should first "feel" the solution, reason internally in that optimised space, then, only when interacting with a human should it convert all that into actual tokens/language.
dragonwriter 40 minutes ago [-]
> This is just temporary though, right?
Well, everything—even human life on Earth—is just temporary, but RAM supply lagging centralized-AI-driven demand increases continuing to squeeze the consumer market may not be a short term phenomenon.
> LLMs seems quite simple, maybe we'll be able to print/assemble at home our own chips with the desired models/weights.
So, the solution to the RAM crunch is “everyone has their own home chip fab and deals with the raw material supply and hazardous waste disposal”?
I...don't imagine so.
XCSme 27 minutes ago [-]
Hazardous waste?
My idea was more like you get some pre-made chips, that you can maybe assemble together configure at home with your desired models.
Maybe each one of this chip is a layer, so you can stack as many layers as you want.
sixothree 44 minutes ago [-]
I fear the future of local models will be controlled by governments. I feel like some time soon there's going to be a crackdown on what is available to download, what is hostable, and what is "acceptable". I partially suspect it has something to do with why 128 GB seems to be the most you can currently purchase for a single machine, despite the price.
matheusmoreira 39 minutes ago [-]
And we push ever closer to the edge of cyberpunk. Psst, interested in some black market heretical Mythos weights?
jrflo 2 hours ago [-]
Let's not forget the Bitter Lesson. Small models sound really nice but at some point you're just fighting the laws of information theory. Efficiency gains on the small model side are nice, but efficiency gains + giant model tends to be even better...
boredatoms 9 minutes ago [-]
How many bits of information are in a real brain?
TeMPOraL 1 hours ago [-]
[dead]
schainks 3 hours ago [-]
I am literally betting my company on this being true.
itsmeduncan 2 hours ago [-]
Me too. I think there are a few waves we can ride here. Let's collaborate?
jmward01 2 hours ago [-]
What company? I am 100% focused on this as a concept in my own internal research.
oblio 2 hours ago [-]
It's a bad bet, historically.
I'm having an extremely hard time thinking of companies that have prospered due to software optimization. Most of them were swept away by hardware advances, instead.
jmward01 2 hours ago [-]
The 1980's US car industry comes to mind. Nearly wiped out because they refused to make efficient vehicles. SpaceX is arguably showing how a rethink towards efficient can take over an entire industry. I am sure there are strong examples in software as well but they aren't coming to mind.
I think when successful, optimization really just means 'finally built right' and people forget the ridiculously inefficient ways before.
sgc 1 hours ago [-]
I would argue about 30% of Apple's success was just from not being slow annoying bloatware. I could never stomach it, but I know a lot of people who basically rage-quit Windows for the Apple ecosystem. I did the same, but for Linux.
somethingweird 2 hours ago [-]
Many of the current internet titans started by making things more efficient and accessible. Google for search, Facebook for connecting to people online, Microsoft for working with PCs at a reasonable price, Amazon for buying online as well as AWS. There are examples in other industries as well, Toyota is famous for it for example. There are probably counter examples but efficiency gains can be a huge deciding factor making companies successful.
bravura 2 hours ago [-]
You just listed a bunch of 0 to 1 companies, not 1 to 10 companies.
They weren’t quantitatively better than previous companies. They were qualitatively better.
cootsnuck 2 hours ago [-]
I think finding significant efficiency gains with LLMs and the like may lead to qualitatively better products. Looking at people's experiences to DSV4F makes me believe that even more than before too.
I don't think people are realizing that speed can allow for categorically different user experiences that are more than just "worse than frontier capabilities but faster".
hgoel 2 hours ago [-]
The headroom for hardware advances is a lot lower now than it has been for most of the industry's existence, when Moore's law held strong. Now we find ourselves limited by cost, physics, fab capacity, and complexity of spinning up more fab capacity.
cootsnuck 2 hours ago [-]
Betting on innovation continuing to figure out ways to squeeze more out of less has historically been the right move. Look at Apple.
And I'd argue "hardware advances" are more proof of optimization.
polymer8563 2 hours ago [-]
IBM wants a word
oblio 1 hours ago [-]
On what? OS/2 didn't fail due to lacking software optimization.
whimsicalism 1 hours ago [-]
I think the path of least resistance will end up being the cheapest and that is scaling up the parameters a ridiculous amount until you get highly capable models that can develop/distill/design the RAM efficient models. Going straight for low param is foolish and just a cope by smaller labs because they don't have the compute/talent to train the large ones.
This is 100% true for pretrains, likely true for RL as well although maybe there is some benefit to smaller activated params there. There is of course 0 benefit to small dense models relative to large sparse ones that are equally as memory efficient if you have enough computers.
Many on HN are in deep denial about this imo.
NBJack 3 hours ago [-]
I honestly hope to see this across all applications, games, services, operating systems, etc. We've been in a period of wasteful RAM usage for over a decade. Constraints, whatever their origin, can be a good thing.
pjmlp 2 hours ago [-]
Same here, back to when algorithms and data structures mattered.
oblio 2 hours ago [-]
If China makes half decent RAM I would bet more on things like 128GM of RAM being the default on low spec laptops 10 years from now.
While I do love optimized software, the hardware side, especially for PCs, has been stagnating for way too long. At least now we have a valid use case for doubling available RAM every 2-3 years again.
I had a reasonably beefy Lenovo consumer line laptop that I bought in 2011, 8GBs of RAM. Its screen hinge broke and I couldn't repair it but I'm fairly sure it was otherwise still usable in 2023-24, once the HDD was replaced with an SSD. I think even now entry level laptops are sold with 8GB of RAM.
By comparison a PC from 2000 was utterly unusable in 2012-13.
KaiMagnus 2 hours ago [-]
I count on a 128GB baseline in 10 years. Beyond the current atmosphere of despair, I really want to see what Apple especially is cooking. Local AI is right up their alley and the current scarcity is unacceptable for them in so many ways.
I got the feeling laptops gonna feel very different in 2036.
mortsnort 46 minutes ago [-]
Really? I feel like because nobody has RAM they're being pushed to the cloud frontier models. If we could all have our own 64GB+ VRAM GPUs, I feel like the open weight model scene would be even stronger.
thehamkercat 4 hours ago [-]
> NeMo Switchyard, an open source library for smart routing
> When deployed, NeMo Switchyard can intelligently direct each request to the most capable and suitable model for the job
How do routers like this handle prompt caching when you send the second request?
Sticky models per session? but then the second message of that session won't be sent to a suitable model, and will only be sent to the same model as previous one.
eli 4 hours ago [-]
I've seen ones that are configurable to pick a trade off point between lower cost (cache stickiness) and routing performance (best model for that turn).
But yeah I'm skeptical all this overhead is worth it.
embedding-shape 4 hours ago [-]
The repo is probably a better entrypoint to it, bit more concise description than the press releases: https://github.com/NVIDIA-NeMo/Switchyard (Notably: "Experimental software. Not for production use."). Unclear if they actually want you to deploy it or not, press release says yes, README says no, do with that what you will.
Doesn't seem to mention "cache" in the README nor the docs, but the code has mentions of it (https://github.com/search?q=repo%3ANVIDIA-NeMo%2FSwitchyard+...), I'm not sure what their thinking is there. "Good luck" essentially? Seems to be per-provider at best, but weird position for a routing library to take.
thehamkercat 4 hours ago [-]
I personally think it's snake-oil marketing with all these smart-model-routing products/projects
prompt-cache won't work with these
try-working 4 hours ago [-]
To keep it simple, forget about routers and imagine you're in Cursor using GPT for a while, reaching a cache of says 200k.
You decide to switch to DeepSeek in the same session via the model picker, and continue as usual. What happens is that the cache for DeepSeek is created with the 200k + the incremental message. After this, cache can be kept warm for both models; two instances of the cache exists, one for GPT and one for DS.
You switch back to GPT. The whole session is sent to the model with the 200k original from GPT and the incremental messages you sent to DS. The 200k is read from cache and the incrementals are new, and then added to the cache.
Let's say every second message you switch between GPT and DS; cache was 200k and each incremental message is 1k. If you kept going with only GPT, cache hit rate would be 200k/(200k+1k) = 99.5%. When you switch between two models with warm cache, hit rate instead becomes 200k/(200k+2k) = 99%.
Model routers work the same way. Keep the cache warm, replicate it in two places. For this reason, when you set up your model pool for routing, you want to keep the model pool small and differentiated.
note: edited to keep the answer to the below message clearer
hedgehog 4 hours ago [-]
To elaborate, because I don't think some of the people reading this understand the reason, typically a lot or most of the cost in "agentic" API usage is cached read + generation. Cached read costs scale with turn count, which multi-model switching doesn't increase, and of course generation gets cheaper if you do some of it with a cheaper model. When you switch models the "catching up" batch of messages is just a single prefill and then that goes into cache. You don't even need to have the same chat history across models so long as the view from each model's perspective looks like a series of appends.
The main problem with model routing in my experience is that to work well the router needs to be pretty strong, maybe even moreso than any of the actual models in service. There are probably clever solutions to this but I haven't seen any that look better than just using sub-agents.
richwater 3 hours ago [-]
> which multi-model switching doesn't increase
Given model A with cache C(a) and model B with C(b)
Isn't this not true because the moment you switch models from A to B, you need to provide C(b) the latest conversation diff since C(b) last updated, say many turns ago?
WASDx 3 hours ago [-]
If you take 10 turns with a model A, it has to read the cache 10 times and write a lot of tokens (the expensive part). Switching to model B is just prefilling the diff + your new message, which is still just one turn. So total number of turns does not increase for a long session even with many switches.
I didn't understand this before reading the sibling comments so I'm not sure I got it fully right but I think the total cost becomes like this:
* Input tokens: Pay for both models
* Output tokens: Pay for the model that generates
* Cached tokens: Pay per turn, so in total a weighted average over both models?
Since output tokens are the most expensive, I can see how this is an overall win for many use cases as benchmarks also show. The hard part is routing correctly.
hedgehog 2 hours ago [-]
[dead]
thehamkercat 4 hours ago [-]
Can you explain how does it work? like how is the previous K/V cache used when you switch to another model?
Source?
hedgehog 3 hours ago [-]
See sibling answer but essentially the effectiveness of cache is not diminished by having a separate one per model (relative to the win of doing more turns and generation with a cheaper model).
try-working 3 hours ago [-]
edit: updated the answer above to be more qualitative instead
quinncom 2 hours ago [-]
Caching should be possible as long as all the models use the same shared cache. The models don't even need to be running on the same server if the shared cache is distributed.
I have a feeling people reading this are thinking that a model router would be used to route between different providers. And in that case, a shared cache would be impossible, although some caching would still be effective. I think, ideally, a router like this is in front of a set of models hosted in one place.
IanCal 2 hours ago [-]
How do caches work across models? I would have thought that was very model specific - if not I’ve really misunderstood what’s getting cached.
armanckeser 2 hours ago [-]
I am not sure the author of the comment you are replying to understands that LLM systems have prompt caches
amluto 2 hours ago [-]
Huh?
Prompt caching isn’t about caching the literal text of the prompt. It’s about caching the result of running prefill on the prompt (or, equivalently, the result of generating the prompt one token at a time by autoregressive inference, or some combination of the above in the case of speculative decoding). This is often called the “KV” cache, and it is very model-specific.
average_bloke 3 hours ago [-]
I would like to propose something:
- problem: massive deluge of information because of AI
- solution: human beings should adopt a minimalist style of communicating in writing.
- e.g. this entire website page can be ten bullet points.
throwatdem12311 49 minutes ago [-]
Why use more words when few do trick?
dofm 18 minutes ago [-]
nods
fooker 2 hours ago [-]
k
stavros 2 hours ago [-]
While I agree with the spirit, I don't think the solution to bad prose is slightly less bad prose. We can write good prose instead.
sp1nningaway 30 minutes ago [-]
hell yeah
encrux 3 hours ago [-]
In my opinion: the only way forward is zero-knowledge-proof authenticated social media.
We can’t have legitimate debate if we have to assume a few bad actors are cloning their voice by the thousands, poisoning debate.
If we can pin one account to a real person, we won’t get rid of LLM-content and misinformation, but at least we can hold them accountable.
ttoinou 3 hours ago [-]
Is the network based on trust and peer to peer confirmation of private keys from who you know in real life that you validated isn’t a robot ?
Or do you have something else in mind ?
kubelsmieci 3 hours ago [-]
> We can’t have legitimate debate
I'm not sure people really want that
npunt 3 hours ago [-]
[dead]
SMAAART 3 hours ago [-]
[flagged]
marsven_422 3 hours ago [-]
[dead]
docheinestages 2 hours ago [-]
They conveniently decided not to include the Qwen range of models in the Artificial Analysis graph, except the out-of-league Max variant. At least be brave and honest.
mark_l_watson 1 hours ago [-]
I love the wave of new small model releases. Pleasantly surprising that an NVIDIA model runs so well on Apple Silicon using MLX! I was using nemotron-3.5-lightning:30b-mlx with OpenCode on my old (cheap) Mac this morning and no bad experiences except for running slowly.
Nemotron 3.5 Lightning runs on how little GPU vram? Can q4 run on 16gb?
sleepyeldrazi 2 hours ago [-]
Not by the looks of it, but it got me thinking, currently in the middle of Level1Techs coverage on the model and switchyard and he mentions "how easy it is to customize it". Fully admitting that I haven't yet read the docs, my issue with that is "we can train LORAs for 35B as well, why use this (according to benchmarks) worse model for customization instead of a slightly bigger better one?"
Assuming I eat my words after going through the docs and this is actually a more efficient model / loras adapt better, I don't see as much value in it as is, as a REAP of it (remove least-important experts, domain-locked tests show ~98% retained accuracy) to something like 20B-A3B (rouhgly matching gpt oss, which while a good model, is outdated knowledge-wise and not as good with tool in my xp).
Having a 20B-A3B model at q4 that has a lora to be your local orchestrator (delegating coding to server/cloud models) and ci/cd runner does start sounding like an appealing proposition to me, as that would fit in 16gb vram easily (fitting many consumer gpus and 24gb macs).
Muse Glimmer 30B seems to be on par with Qwen 3.6 27B (4 months old)
but
Qwen 3.8 27B is dropping this week...
XCSme 3 hours ago [-]
Yes, I was surprised to see doing it as well as Qwen 3.7 27b.
Even though that model is already "old", qwen was way ahead everyone else in that size category before this Meta model.
Also, probably for non-Chinese usage, using a non-Chinese model might lead to better results.
eli 4 hours ago [-]
The top 4 models on that site are all variants of Gemini Flash? That does not match my experience at all.
XCSme 4 hours ago [-]
I should add a F.a.q. for this question.
The suite is across many categories, not only coding, and most of the tasks are low-horizon (or what the opposite of long-horizon is), where the max thinking time is around 10 minutes.
Gemini models are really smart, unfortunately they don't play well with any harness, so hard to use in practice.
But try them out for one-shot tasks, they are really good. Don't use them for coding in a harness, but you can ask them to generate code/planning (still, for coding only other models are indeed recommended).
markasoftware 2 hours ago [-]
yep, the person you're responding to created the benchmark and is using HN comments as advertisement.
XCSme 2 hours ago [-]
I did, avertisment is a big word, as I gain nothing from the traffic, I run the website for myself, and some other people find it useful too.
Happy to hear what would make the website more useful.
khimaros 2 hours ago [-]
lightning is sparse, glimmer is dense
XCSme 2 hours ago [-]
Oh, good to know, I just quickly tested and published the results.
I will add model sizes (total/active params) for each model, good point.
Tactical45 4 hours ago [-]
At what cost difference?
XCSme 4 hours ago [-]
I don't think it matters, if it's for local/on-device usage.
The cost is similar vram footprint I guess (?)
sleepyeldrazi 1 hours ago [-]
loading the model would be similar vram footprint, correct, however the size of KV is based on 'active' params, not total params. So while at 1k ctx both will be in the same ballpark vram footprint-wise, at 100k the story will be very different. 27B at q4 kv for 256k takes ~8gb, while 35B at q4 kv around ~3.5gb, so at full precision kv those would be ~32gb and ~14gb (all ballparks, if you want exact numbers, its not hard to test).
As for the "cost", here i think the interesting arguments are around speed vs accuracy/"getting the job done", not literal $ cost per token.
XCSme 1 hours ago [-]
I am asking mostly for running on a 3090.
I think the tps difference between them (both fitting in vram) won't be more than 2x in practice.
I would happily take 20tps over 40tps, if the model gets 3x more correct answers.
sleepyeldrazi 48 minutes ago [-]
Speed is (for the most part) active-parameter based, so a 30B-A3B model is roughly 10x the speed of a dense 30B (realistically closer to 8x) in the case when both fit. That's the proposition of MoE and why everyone is trying to make massive models with very few active params, so that they are still fast while having access to a lot of knowledge (at the cost of reasoning, as reasoning ability 'for the most part' comes from active params).
You can test this by running this nemo or 35B on the 3090. I have and its very fun (but sadly a worse model than 27B, so I usually keep 27B on my 3090)
XCSme 36 minutes ago [-]
I remember running both qwen 30b-a3b and 27b on my 3090, and on the initial test, the 27b was only like 2x slower.
rllearneratwork 2 hours ago [-]
and Glimmer has 10x active params of Lightning. Meaning ~ 10 slower on same HW
XCSme 2 hours ago [-]
Is that the case?
If the entire model fits in vram, won't the tps be comparable?
option 47 minutes ago [-]
autoregressive generation is memory bandwidth bounded
XCSme 30 minutes ago [-]
So, on a crude calculation, for a 3090 with 936.2 GB/s, a model that has 20 GB of active params would run at 45tps and one with 3GB active params at 300tps?
In practice, I don't think I saw over 100tps on a 3090, for a local 20-30B model, be it MoE or not.
I'm looking forward to seeing what types of new things people create over the coming years once there is less obsession with massive unwieldy LLMs. I think the incentives are just too strong to ignore.
With the benefit of LLMs already being proven, in a couple of years we will have vastly better hardware for inference I guess.
I feel like now hardware is stagnating a bit, because the software side has moved too fast for the hardware to catch up. Once we settle on some good, optimal software architecture for the models, dedicated hardware will easily increase throughout by 10x or 100x, for a fraction of the cost.
LLMs seems quite simple, maybe we'll be able to print/assemble at home our own chips with the desired models/weights.
Maybe we'll have model weights being shared like game cartridges.
This has a lot to do with how MCTS works BTW. The current best path is often only the current best path because a lot of investment has been sunk into it. If you were to put equal resources into a different path you may find that it was actually far better. It is just that the early rollouts favored the other 'best path' so you sunk a lot of resources into that one. We are very early in our exploration of LLM architecture. I highly doubt we are anywhere near the best path right now.
The diffusion models are interesting, but those also seem hacky.
I think the next form of AIs will be simpler and more abstract.
The building blocks of our brain don't have the notion of a "token" embed into them, it's lower level that that.
I think first step is to find a better way to represent information.
LLMs shouldn't "compute" stuff using language tokens, but some other, more efficient logical mechanisms. LLMs should first "feel" the solution, reason internally in that optimised space, then, only when interacting with a human should it convert all that into actual tokens/language.
Well, everything—even human life on Earth—is just temporary, but RAM supply lagging centralized-AI-driven demand increases continuing to squeeze the consumer market may not be a short term phenomenon.
> LLMs seems quite simple, maybe we'll be able to print/assemble at home our own chips with the desired models/weights.
So, the solution to the RAM crunch is “everyone has their own home chip fab and deals with the raw material supply and hazardous waste disposal”?
I...don't imagine so.
My idea was more like you get some pre-made chips, that you can maybe assemble together configure at home with your desired models.
Maybe each one of this chip is a layer, so you can stack as many layers as you want.
I'm having an extremely hard time thinking of companies that have prospered due to software optimization. Most of them were swept away by hardware advances, instead.
I think when successful, optimization really just means 'finally built right' and people forget the ridiculously inefficient ways before.
They weren’t quantitatively better than previous companies. They were qualitatively better.
I don't think people are realizing that speed can allow for categorically different user experiences that are more than just "worse than frontier capabilities but faster".
And I'd argue "hardware advances" are more proof of optimization.
This is 100% true for pretrains, likely true for RL as well although maybe there is some benefit to smaller activated params there. There is of course 0 benefit to small dense models relative to large sparse ones that are equally as memory efficient if you have enough computers.
Many on HN are in deep denial about this imo.
While I do love optimized software, the hardware side, especially for PCs, has been stagnating for way too long. At least now we have a valid use case for doubling available RAM every 2-3 years again.
I had a reasonably beefy Lenovo consumer line laptop that I bought in 2011, 8GBs of RAM. Its screen hinge broke and I couldn't repair it but I'm fairly sure it was otherwise still usable in 2023-24, once the HDD was replaced with an SSD. I think even now entry level laptops are sold with 8GB of RAM.
By comparison a PC from 2000 was utterly unusable in 2012-13.
I got the feeling laptops gonna feel very different in 2036.
> When deployed, NeMo Switchyard can intelligently direct each request to the most capable and suitable model for the job
How do routers like this handle prompt caching when you send the second request?
Sticky models per session? but then the second message of that session won't be sent to a suitable model, and will only be sent to the same model as previous one.
But yeah I'm skeptical all this overhead is worth it.
Doesn't seem to mention "cache" in the README nor the docs, but the code has mentions of it (https://github.com/search?q=repo%3ANVIDIA-NeMo%2FSwitchyard+...), I'm not sure what their thinking is there. "Good luck" essentially? Seems to be per-provider at best, but weird position for a routing library to take.
prompt-cache won't work with these
You decide to switch to DeepSeek in the same session via the model picker, and continue as usual. What happens is that the cache for DeepSeek is created with the 200k + the incremental message. After this, cache can be kept warm for both models; two instances of the cache exists, one for GPT and one for DS.
You switch back to GPT. The whole session is sent to the model with the 200k original from GPT and the incremental messages you sent to DS. The 200k is read from cache and the incrementals are new, and then added to the cache.
Let's say every second message you switch between GPT and DS; cache was 200k and each incremental message is 1k. If you kept going with only GPT, cache hit rate would be 200k/(200k+1k) = 99.5%. When you switch between two models with warm cache, hit rate instead becomes 200k/(200k+2k) = 99%.
Model routers work the same way. Keep the cache warm, replicate it in two places. For this reason, when you set up your model pool for routing, you want to keep the model pool small and differentiated.
First principles of model routing: https://try.works/first-principles-of-model-routing
role-model router and protocol: https://github.com/try-works/role-model
note: edited to keep the answer to the below message clearer
The main problem with model routing in my experience is that to work well the router needs to be pretty strong, maybe even moreso than any of the actual models in service. There are probably clever solutions to this but I haven't seen any that look better than just using sub-agents.
Given model A with cache C(a) and model B with C(b)
Isn't this not true because the moment you switch models from A to B, you need to provide C(b) the latest conversation diff since C(b) last updated, say many turns ago?
I didn't understand this before reading the sibling comments so I'm not sure I got it fully right but I think the total cost becomes like this:
* Input tokens: Pay for both models * Output tokens: Pay for the model that generates * Cached tokens: Pay per turn, so in total a weighted average over both models?
Since output tokens are the most expensive, I can see how this is an overall win for many use cases as benchmarks also show. The hard part is routing correctly.
Source?
I have a feeling people reading this are thinking that a model router would be used to route between different providers. And in that case, a shared cache would be impossible, although some caching would still be effective. I think, ideally, a router like this is in front of a set of models hosted in one place.
Prompt caching isn’t about caching the literal text of the prompt. It’s about caching the result of running prefill on the prompt (or, equivalently, the result of generating the prompt one token at a time by autoregressive inference, or some combination of the above in the case of speculative decoding). This is often called the “KV” cache, and it is very model-specific.
- problem: massive deluge of information because of AI
- solution: human beings should adopt a minimalist style of communicating in writing.
- e.g. this entire website page can be ten bullet points.
We can’t have legitimate debate if we have to assume a few bad actors are cloning their voice by the thousands, poisoning debate.
If we can pin one account to a real person, we won’t get rid of LLM-content and misinformation, but at least we can hold them accountable.
Or do you have something else in mind ?
I'm not sure people really want that
Assuming I eat my words after going through the docs and this is actually a more efficient model / loras adapt better, I don't see as much value in it as is, as a REAP of it (remove least-important experts, domain-locked tests show ~98% retained accuracy) to something like 20B-A3B (rouhgly matching gpt oss, which while a good model, is outdated knowledge-wise and not as good with tool in my xp).
Having a 20B-A3B model at q4 that has a lora to be your local orchestrator (delegating coding to server/cloud models) and ci/cd runner does start sounding like an appealing proposition to me, as that would fit in 16gb vram easily (fitting many consumer gpus and 24gb macs).
https://aibenchy.com/compare/meta-muse-glimmer-30b-xhigh/nvi...
but
Qwen 3.8 27B is dropping this week...
Even though that model is already "old", qwen was way ahead everyone else in that size category before this Meta model.
Also, probably for non-Chinese usage, using a non-Chinese model might lead to better results.
The suite is across many categories, not only coding, and most of the tasks are low-horizon (or what the opposite of long-horizon is), where the max thinking time is around 10 minutes.
Gemini models are really smart, unfortunately they don't play well with any harness, so hard to use in practice.
But try them out for one-shot tasks, they are really good. Don't use them for coding in a harness, but you can ask them to generate code/planning (still, for coding only other models are indeed recommended).
Happy to hear what would make the website more useful.
I will add model sizes (total/active params) for each model, good point.
The cost is similar vram footprint I guess (?)
As for the "cost", here i think the interesting arguments are around speed vs accuracy/"getting the job done", not literal $ cost per token.
I think the tps difference between them (both fitting in vram) won't be more than 2x in practice.
I would happily take 20tps over 40tps, if the model gets 3x more correct answers.
You can test this by running this nemo or 35B on the 3090. I have and its very fun (but sadly a worse model than 27B, so I usually keep 27B on my 3090)
If the entire model fits in vram, won't the tps be comparable?
In practice, I don't think I saw over 100tps on a 3090, for a local 20-30B model, be it MoE or not.