You mean Mercury 2, by Inception: <a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;inception&#x2F;mercury-2" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;inception&#x2F;mercury-2</a>

You are right, people who downvoted you are just ignorant.

That&#x27;s completely different. That&#x27;s like saying you want to compare the Nvidia 5090 GPU to the latest Call of Duty.

Not wrong, but I think it&#x27;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&#x27;s a different problem setup.

I mean I guess but the diffusion objective and the ability to do simultaneous decode both dictate pretty different architectures in practice.

Mamba-3 is an architecture while diffusion is, I believe, a type of objective. So these are not mutually exclusive and therefore not comparable.

I&#x27;m looking forward to comparing this to Inception 2 (the text diffusion model) which in my experience is very fast and reasonably high quality.

Can 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.

Because something linear like Mamba doesn&#x27;t perform as well; so you&#x27;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&#x27;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 <a href="https:&#x2F;&#x2F;magazine.sebastianraschka.com&#x2F;i&#x2F;168650848&#x2F;18-nemotron-3-nano-and-super" rel="nofollow">https:&#x2F;&#x2F;magazine.sebastianraschka.com&#x2F;i&#x2F;168650848&#x2F;18-nemotro...</a> for some discussion on this architecture

They did do that, 2 years ago. The problems are that 1) mamba makes accuracy worse as context size grows, 2) Nvidia GPUs are designed for transformers, and 3) all the software out there is also designed for transformers. It&#x27;s still useful in some applications but it doesn&#x27;t beat regular transformers if you have the gear

Probably 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.

We kinda do do this with hybrid mamba transformers

Linear time complexity models are bad at in-context retrieval, which limits their performance on various tasks, so a pure linear model isn&#x27;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.

Is 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

Well, 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&#x2F;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&#x27;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.

&gt; Everyone groups all the requests into a batch, and the GPU computes them together.You&#x27;re only saving on fetching read-only parameters, and not even on that if you&#x27;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&#x27;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.

Focusing on needs of providers isn&#x27;t a very good long term strategy if you believe compute will eventually move to self hosted and on premises solutions where large batch sizes aren&#x27;t needed.

Why would execs need latency?I can see it for engineering - coding with slow ai is painful

Throughput is indeed king for the standard-tier mindshare-capture play. But there are many who would pay multiple times the current cost for agentic systems for engineers and executives, if it meant a meaningful reduction in latency. The economics could work extremely well.

Local has a batch size of 1. If you are already memory bound then you leave compute on the table. Why not use it?Not sure they target local though…

Well, Deepseek batch sizes are something like 8192, so 128 isn&#x27;t much.<a href="https:&#x2F;&#x2F;arxiv.org&#x2F;html&#x2F;2412.19437v1" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;html&#x2F;2412.19437v1</a> &quot;the batch size per expert is relatively small (usually within 256 tokens)&quot;

Their latency measurements comparing Mamba-2 and Mamba-3 are done with a batch size of 128. It doesn&#x27;t seem like Mamba-2 was compute-bound even at that batch size.

I&#x27;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&#x27;t leaving GPU cores idle normally during inference.

Probably constrained by training resources. It&#x27;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.

This 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?

I&#x27;m glad I clicked through bc I thought the article was about Mamba, the package manager I associate with Python (similar to conda).<a href="https:&#x2F;&#x2F;github.com&#x2F;mamba-org&#x2F;mamba" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;mamba-org&#x2F;mamba</a>

More here <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47423208">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47423208</a><a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2603.15569" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2603.15569</a>

I&#x27;m looking forward to the fifth iteration of this model.

They don&#x27;t say anything about dropping training speed.

This 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&#x27;s a nice opening as it is imo

Agreed. What you wrote was probably the input, what we see is the LLM output with the directive to &quot;make us sound smart, put gratuitous em-dash&quot;

Of 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.

The first sentence basically does though, no?

The blog is technical, technical terms in the TL;DR seems relevant to me.

Yes, it is written for a specific audience.That is not a reason for snark.As other commenters have noted, it’s well written.

&gt; I don&#x27;t get the downvotesBecause the blog post is a technical one and the intro contains very common jargon, and the proposed alternative was wrong.

I don&#x27;t get the downvotes, as I had trouble understanding the intro as well. It seems it was written for a very specific audience.

Found the guy who made the Windows error messages say “Your computer did an oopsie :(“ instead of including any useful information.

Why would the simpler version be better for a technical audience?

I 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.

&gt; 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.