I don’t understand why this lab is allergic to providing details on what they actually made, especially when Chinese labs are more than willing to share architectural specs/code/kernels (eg NSA/FSA, RAMBa, HISA, DSA LightningIndexer, etc). I don’t doubt that they’ve done something here, but the lack of details makes me default not trust this, particularly when this is the second time that they’ve released a “technical report” that just waxes poetic about the concept.
Business wise, it would make sense to hold off on details till they're at least ready to serve. Look at what happened with Open AI and reasoning models. Everyone struggled with getting RL to work with LLMs for a good while. Open AI figured it out, and a few months later everyone had their prototypes out in short order. Don't forget who these labs employ. They're some of the brightest people around. Sub-q aren't really in a position for that lol. If they'd shared details at the first announcement for instance, the big labs might have had something out by now while they're still pulling resources to scale and then what ?
You don't understand why the thing their entire company is valued upon is...not being given away freely? They literally are taking an open source model and then adapting it with this technique. If they disclose it, the frontier labs will immediately copy it and outperform them.
My guess is that they're angling for an acquisition.
This one's interesting, and I think the next frontier for LLMs should really just be, how can we get something like Opus 4.6 to cost drastically less, for the same output? I say 4.6 because from 4.6 onwards it's been pretty darn good, at least for me, always feels like every model upgrade someone hates it, heck even 4.5 was fine.
Yes - I want that and dramatically faster. Newer models don't seem to need any more or less guidance and iteration, so let's make the time-to-wrong-answer as short as possible.
I'm not as crazy about speed as long as it's reasonably as "quick" as Opus. Which is faster than most developers can spit out code. I do get annoyed with Claude Code because it looks like it chooses to be as slow as possible, but maybe that's by design so its not pounding their backend every milisecond? Would probably be bad.
Local inference is insanely fast on my M4 Pro MBP though, so I can understand where you're coming from, but I don't need it too much faster. I still need time to review, test, review and provide feedback to the model. Fast is okay I guess for true vibe coding.
I just don't want to have to have a pipeline going in order to fully occupy my time. I don't want to wait on the model to review the prompt, read the parts of the codebase indicated, do its own research in the codebase and documentation, plan, run agents ... actually write the code and NOW I can start reading it and reviewing it. That means I either need to run a lot of operations in parallel so that I always have something to do and the agent(s) are highly utilized or I'm writing something on my own that I keep getting that keeps getting interrupted. It's the constant context switching that kills me. I want to work on one problem at a time and really focus on it - even if I'm not writing every line myself.
> SubQ 1.1 Small scores near-perfect at 1M, 2M, 6M, and 12M tokens. The model was trained predominantly at 1M tokens yet the retrieval held near perfectly at 12x that length, despite compressing attention to just 0.13% of relationships. This generalization is a direct consequence of SSA routing attention based on content relevance rather than fixed positional patterns.
If the results persists from 1M to 12M, why not 24M or 48M? Sounds almost too good to be true.
With back of the napkin math from inside my head, that'd be like 0.5/1 million LOC, depending on language/code density, could just fold the entire codebase into one prompt if it's a small one, that'd be neat :)
What kind of hardware would be needed to serve an instance with the full 12m context? And what kind of speeds can one expwct at those extremes at 10m+?
They've done multiple "evaluations" by third parties, but still, it seems that they aren't being fully transparent. I think the approach is quite interesting and novel, but this feels like deja vu.
I get why they aren't disclosing all the details, but it seems more hype-train-esque to me for this moment. I don't disagree that this could be big.
FlashAttention-2 is not used anymore for at least 2y.
This architecture would have been a massive improvement 3 years ago, but it is a ~solved~ problem IMO.
My guess is that they're angling for an acquisition.
Local inference is insanely fast on my M4 Pro MBP though, so I can understand where you're coming from, but I don't need it too much faster. I still need time to review, test, review and provide feedback to the model. Fast is okay I guess for true vibe coding.
> At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2.
Awesome stuff. Solving context at the model architecture layer rather than trying to bolt on extra memory is the right direction IMO.
If the results persists from 1M to 12M, why not 24M or 48M? Sounds almost too good to be true.
With back of the napkin math from inside my head, that'd be like 0.5/1 million LOC, depending on language/code density, could just fold the entire codebase into one prompt if it's a small one, that'd be neat :)
I get why they aren't disclosing all the details, but it seems more hype-train-esque to me for this moment. I don't disagree that this could be big.