7 ms·
Mamba-3
- robofanatic 7mo ago> Mamba-3 is a new state space model (SSM) designed with inference efficiency as the primary goal — a departure from Mamba-2, which optimized for training speed. The key upgrades are a more expressive recurrence formula, complex-valued state tracking, and a MIMO (multi-input, multi-output) variant that boosts accuracy without slowing down decoding. Why can’t they simply say - Mamba-3 focuses on being faster and more efficient when making predictions, rather than just being fast to train like Mamba-2.
- E-Reverance 7mo agoThe first sentence basically does though, no?
- robofanatic 7mo agoOf course my only objection was the language. LLMs are now old enough to leave the jargon behind and talk in simple easy to understand terms.
- oersted 7mo agoI’d argue the opposite, the terminology is fairly mainstream by now and “inference” has a much more specific sense than “making predictions”.
- esquire_900 7mo agoThis is sort of what their first sentence states? Except your line implies that they are fast in training and inference, they imply they are focusing on inference and are dropping training speed for it. It's a nice opening as it is imo
- cubefox 7mo agoThey don't say anything about dropping training speed.
- arendtio 7mo agoI don't get the downvotes, as I had trouble understanding the intro as well. It seems it was written for a very specific audience.
- qeternity 7mo agoYes, it is written for a specific audience. That is not a reason for snark. As other commenters have noted, it’s well written.
- magicalhippo 7mo ago> I don't get the downvotes Because the blog post is a technical one and the intro contains very common jargon, and the proposed alternative was wrong.
- mufasachan 7mo agoThe blog is technical, technical terms in the TL;DR seems relevant to me.
- camillomiller 7mo agoI don’t know why you’re being downvoted. As a longtime editor your version is immensely better. Looks like the original was probably not human-written.
- stavros 7mo agoWhy would the simpler version be better for a technical audience?
- renewiltord 7mo agoFound the guy who made the Windows error messages say “Your computer did an oopsie :(“ instead of including any useful information.
- i000 7mo agoAgreed. What you wrote was probably the input, what we see is the LLM output with the directive to "make us sound smart, put gratuitous em-dash"
- nl 7mo agoI'm looking forward to comparing this to Inception 2 (the text diffusion model) which in my experience is very fast and reasonably high quality.
- cubefox 7mo agoMamba-3 is an architecture while diffusion is, I believe, a type of objective. So these are not mutually exclusive and therefore not comparable.
- gyrovagueGeist 7mo agoNot wrong, but I think it's more accurate to say: Mamba is an architecture for the middle layers of the network (the trunk) which assumes decoding takes place through an autoregressive sequence (popping out tokens in order). This is the SSM they talk about. Diffusion is an alternative to the autoregressive approach where decoding takes place through iterative refinement on a batch of tokens (instead of one at a time processing and locking each one in only looking forward). This can require different architectures for the trunk, the output heads, and modifications to the objective to make the whole thing trainable. Could mamba like ideas be useful in diffusion networks...maybe but it's a different problem setup.
- cubefox 7mo agoLinear architectures are at least heavily used in image diffusion models. More so in fact than in language models.
- joefourier 7mo agoMamba doesn't assume auto-regressive decoding, and you can use absolutely use it for diffusion, or pretty much any other common objective. Same with a conventional transformer. For a discrete diffusion language model, the output head is essentially the same as an autoregressive one. But yes, the training/objective/inference setup is different.
- nl 7mo agoI mean I guess but the diffusion objective and the ability to do simultaneous decode both dictate pretty different architectures in practice.
- jychang 7mo agoI'm not sure that I buy their conclusion that more compute during inference is good. Yes, batch=1 inference is mostly memory bandwidth bound, not GPU compute bound. But no provider does batch=1 inference. Everyone groups all the requests into a batch, and the GPU computes them together. With a fused kernel, that means the GPU streams the tensors from VRAM, and does a bunch of compute on different conversations in the batch, at the same time. If they increase the amount of compute required per token, that just reduces the maximum batch size a GPU can handle. In practice, yes this does mean each GPU can serve less users. Providers aren't leaving GPU cores idle normally during inference.
- zozbot234 7mo ago> Everyone groups all the requests into a batch, and the GPU computes them together. You're only saving on fetching read-only parameters, and not even on that if you're using MoE models where each inference in the batch might require a different expert (unless you rearrange batches so that sharing experts becomes more likely, but that's difficult since experts change per-token or even per-layer). Everything else - KV-cache, activations - gets multiplied by your batch size. You scale both compute and memory pressure by largely the same amount. Yes, GPUs are great at hiding memory fetch latency, but that applies also to n=1 inference.
- jychang 7mo agoWell, the actual inference providers put each expert on its own single GPU. Deepseek explicitly does this. Read-only parameters is also usually the majority of space. Deepseek is 700GB of params. Meanwhile kv cache is small (Deepseek is about 7GB at max context) and ssm/conv1d cache is even smaller- IIRC Qwen 3.5 is 146MB per token regardless of context size. Not sure about how Mamba-3 works, but I suspect read-only parameters are still a significant amount of memory bandwidth. I guess the question isn't whether compute is 1:1 with memory, but rather if you run out of compute before you run out of vram adding more users.
- zozbot234 7mo ago> Well, the actual inference providers put each expert on its own single GPU. Experts are usually chosen on a per-layer basis, not just by token, so I'd think this requires having lots of GPU's to make it worthwhile. You could do it with a single physical GPU by switching expert-layer mixes in a round-robin fashion after the batch for any single expert-layer mix is completed (essentially a refined version of expert offloading). But still, not easy.
- daliliu 7mo ago[dead]
- Havoc 7mo agoIs there a reason we don’t switch halfway through? ie start with a classic LLM and switch to something linear like mamba as context grows
- energy123 7mo agoProbably best achieved by model routing, either an indirection behind the chat UI or an API user does it themselves by calling a different API for long context queries.
- cubefox 7mo agoLinear time complexity models are bad at in-context retrieval, which limits their performance on various tasks, so a pure linear model isn't currently feasible anyway, at least for language models. Instead they recommend mixing linear and attention layers. Presumably this mostly solves the performance problem (at least n benchmarks), but it also means the mixed architecture is no longer linear. It will still be faster and less RAM hungry in long context than a pure transformer though.
- lambda 7mo agoBecause something linear like Mamba doesn't perform as well; so you'd have a performance cliff, where suddenly the model would get more dumb and forget a lot of what was going on. Instead, you can get benefits from both by doing both in parallel. This can let you reduce the size of the O(n^2) attention mechanism, so while it's still quadratic, it reduces the constant quite a bit while still retaining a lot of performance, as the linear context mechanism can work for the tasks its well suited for while allowing attention to play to its strengths. The recent Nemotron 3 Nano and Super models from NVIDIA are hybrid architectures this way, with most of their context layers as Mamba while retaining enough attention to continue to be competitive on the more complex tasks that require the quadratic attention. See https://magazine.sebastianraschka.com/i/168650848/18-nemotron-3-nano-and-super https://magazine.sebastianraschka.com/i/168650848/18-nemotro... for some discussion on this architecture
- 3abiton 7mo agoI am curious of the tradeoff of hybrid approaches, it sounds too good to be true.
- diablevv 7mo ago[flagged]
- anentropic 7mo agoMore here https://news.ycombinator.com/item?id=47423208 https://news.ycombinator.com/item?id=47423208 https://arxiv.org/abs/2603.15569 https://arxiv.org/abs/2603.15569
- jeffhwang 7mo agoI'm glad I clicked through bc I thought the article was about Mamba, the package manager I associate with Python (similar to conda). https://github.com/mamba-org/mamba https://github.com/mamba-org/mamba
- manlymuppet 7mo agoI'm looking forward to the fifth iteration of this model.
- breadsniffer 7mo agoMamboooo no. #5
- roger_ 7mo agoCan anyone explain why Mamba models start with a continuous time SSM (and discretize) vs discrete time? I know the step isn’t fixed, also not sure why that’s important. Is that the only reason? There also seems to be a parameterization advantage too with the continuous formulation.
- fudged71 7mo agoThis is really promising. Are they now going to scale this up to hundreds of billions of parameters? Why stop at 1.5B if they found a potentially SOTA architecture?
- snek_case 7mo agoProbably constrained by training resources. It's much easier to experiment with a smaller architecture. You may need many training runs to figure out hyperparameters for example. If each run needs multiple GPUs for a week the cost adds up quickly. I think it makes a lot of sense to start small.