Suggestion for the maintainers: the comparison table currently lists some pretty old models, Qwen 2.5 14B and Mixtral 8x7B and Llama 3.3 70B.
A lot of people are reporting incredible results with the Qwen 3.5 MoE models on Apple hardware right now (streaming experts - see https://simonwillison.net/2026/Mar/24/streaming-experts/) - it would be great to get some of those models into that table.
For a lot of local workloads, sub-1 tok/s is useless in foreground and perfectly acceptable in background. If the choice is “this crashes” vs “this finishes overnight,” that’s still a meaningful capability jump.
the practical question is whether the read pattern is sequential enough to actually saturate nvme bandwidth or if the attention layer access pattern ends up being random enough to kill throughput. sequential reads on a decent nvme get you 5-7 GB/s, random reads drop to maybe 500 MB/s depending on queue depth.
for a 1T model youd need to stream something like 2TB of weights per forward pass at fp16. even at peak sequential thats 300+ seconds per token which is... not great for interactive use but maybe fine for batch inference where you dont care about latency.
still a cool proof of concept though. the gap between 'can run' and 'runs usefully' is where things get interesting.
Yes, definitely agree. It's more of a POC than a functional use case. However, for many smaller MoE models this method can actually be useful and capable of achieving multiple tokens/sec.
> for a 1T model youd need to stream something like 2TB of weights per forward pass
Isn't this missing the point of MoE models completely? MoE inference is sparse, you only read a small fraction of the weights per layer. You still have a problem of each individual expert-layer being quite small (a few MiBs each give or take) but those reads are large enough for the NVMe.
The MoE point matters here ie sparse activation means you're not reading all 2TB per forward pass, but the access pattern flips from sequential to random which is exactly the worst case for NVMe. Been thinking about this a lot for agent inference workloads where you want consistent latency more than peak throughput.
I'm referencing it as being possible, however I didn't share benchmarks because candidly the performance would be so slow it would only be useful for very specific tasks over long time horizons. The more practical use cases are less flashy but capable of achieving multiple tokens/sec (ie smaller MoE models where not all experts need to be loaded in memory simultaneously)
Still have 4 brand new ones in my storage unit. Just in case these moments.
Joke aside (I do have them tho!), I don't think Optane is that much use (not to mention it is only 256GiB for my unit). It is useful legacy crutch if you have legacy software that is not designed to issue multiple reads / writes in parallel. If you do, it is really not faster than NVMe, especially these modern ones.
It's not about being faster (except for small reads where latency dominates, which is actually relevant when reading a handful of expert-layers immediately after routing), it's the wearout resistance which opens up the possibility of storing KV-cache (including the "linear" KV-cache of recent Qwen, which is not append-only as it was with the pure attention model) and maybe even per-layer activations - though this has the least use given how ephemeral these are.
Yes, their NAND division has been sold, it is now mostly under solidigm. Maybe solidigm could bring it back, but it seems unlikely (given the previous commercial failure).
This is a pretty cool project! Essentially this is like using Swap memory to extend your RAM, but in a 'smart' way so you don't overload the NVMe unnecessarily.
I do wonder in practice how the 'smarts' pan out, because putting a ton of stress on your NVMe during generation is probably not the best choice for it's longevity.
> but in a 'smart' way so you don't overload the NVMe unnecessarily
"overloading NVMe"? What is that about? First time I've heard anything about it.
> because putting a ton of stress on your NVMe during generation
Really shouldn't "stress your NVMe", something is severely wrong if that's happening. I've been hammering my SSDs forever, and while write operations "hurt" the longevity of the flash cells themselves, the controller interface really shouldn't be affected by this at all, unless I'm missing something here.
Even if there was a ton of writing, I'm not sure where NVMe even comes in the picture, write durability is about the flash cells on SSDs, nothing to do with the interface, someone correct me if I'm wrong.
There needs to be something like this from Ollama. At the moment Ollama has a lot of flaws that prevent it from getting great performance. (My understanding is better GPU/CPU splits, etc). But Ollama is the only way to host an LLM and have it switch out on demand. Sigh.
Ollama has very substandard support for mmap at present, which hurts inference with larger models. There are some recent pull requests in flight that should help address this to at least some extent https://github.com/ollama/ollama/pull/14525https://github.com/ollama/ollama/pull/14134https://github.com/ollama/ollama/pull/14864 but progress seems to be stalling. Their support for recent Qwen models seems to also have some bespoke incompatibilities with llama.cpp, which doesn't help matters; it's difficult to test the same model with both.
You do not provide any comparison to llama.cpp with mmap.
You do not explain how any kind of predictor can work for MoE experts.
You do not explain how prediction can even be useful. I can predict the layers used in a dense model (all of them are used in order), but that doesn't help me much. It's still bottlenecked on bandwidth (hint: MoE doesn't change this).
OS paging would be significantly worse here. The kernel's page fault handler is reactive — it doesn't know
you're about to read layer 47's FFN weights, so it can't prefetch. You stall on every fault, wait for the
4KB/16KB page to load, then resume. With 80 layers of dense FFN streaming, that's thousands of cold faults per
token.
What makes this approach faster is that the model's access pattern is completely deterministic during
inference. You know exactly which tensors are needed next because transformer layers execute sequentially. So
you can issue large sequential reads and prefetch the next layer while the current one is computing on Metal.
The OS page cache can't do that — it has no concept of "layer N+1 comes after layer N."
For MoE it's even more stark. The OS would page in all 8 experts on the first token that routes to each one,
then evict them under memory pressure with LRU, which has no idea that expert 3 fires 10x more often than
expert 7. The neuron cache here is basically a domain-specific replacement policy.
Suggestion for the maintainers: the comparison table currently lists some pretty old models, Qwen 2.5 14B and Mixtral 8x7B and Llama 3.3 70B.
A lot of people are reporting incredible results with the Qwen 3.5 MoE models on Apple hardware right now (streaming experts - see https://simonwillison.net/2026/Mar/24/streaming-experts/) - it would be great to get some of those models into that table.
Maybe the 1T parameter Kimi K2.5 too if you can get that to work, see https://twitter.com/seikixtc/status/2036246162936910322 and https://twitter.com/danpacary/status/2036480556045836603
For a lot of local workloads, sub-1 tok/s is useless in foreground and perfectly acceptable in background. If the choice is “this crashes” vs “this finishes overnight,” that’s still a meaningful capability jump.
the practical question is whether the read pattern is sequential enough to actually saturate nvme bandwidth or if the attention layer access pattern ends up being random enough to kill throughput. sequential reads on a decent nvme get you 5-7 GB/s, random reads drop to maybe 500 MB/s depending on queue depth.
for a 1T model youd need to stream something like 2TB of weights per forward pass at fp16. even at peak sequential thats 300+ seconds per token which is... not great for interactive use but maybe fine for batch inference where you dont care about latency.
still a cool proof of concept though. the gap between 'can run' and 'runs usefully' is where things get interesting.
4K random read with a queue depth of 1 on an M1 Max is about 65MB/s.
Yes, definitely agree. It's more of a POC than a functional use case. However, for many smaller MoE models this method can actually be useful and capable of achieving multiple tokens/sec.
> for a 1T model youd need to stream something like 2TB of weights per forward pass
Isn't this missing the point of MoE models completely? MoE inference is sparse, you only read a small fraction of the weights per layer. You still have a problem of each individual expert-layer being quite small (a few MiBs each give or take) but those reads are large enough for the NVMe.
But across a sequence you still have to load most of them.
The MoE point matters here ie sparse activation means you're not reading all 2TB per forward pass, but the access pattern flips from sequential to random which is exactly the worst case for NVMe. Been thinking about this a lot for agent inference workloads where you want consistent latency more than peak throughput.
Where does "1T parameter model" come from? I can only see models with 70B params or less mentioned in the repo.
I'm referencing it as being possible, however I didn't share benchmarks because candidly the performance would be so slow it would only be useful for very specific tasks over long time horizons. The more practical use cases are less flashy but capable of achieving multiple tokens/sec (ie smaller MoE models where not all experts need to be loaded in memory simultaneously)
Yeah title comes from nowhere in the link. No doubt it's possible but all that matters is speed and we learn nothing of that here...
Intel Optane rolling in its grave.
Memristors are also missing in this AI hype even when they were around the corner 10 years back.
Still have 4 brand new ones in my storage unit. Just in case these moments.
Joke aside (I do have them tho!), I don't think Optane is that much use (not to mention it is only 256GiB for my unit). It is useful legacy crutch if you have legacy software that is not designed to issue multiple reads / writes in parallel. If you do, it is really not faster than NVMe, especially these modern ones.
It's not about being faster (except for small reads where latency dominates, which is actually relevant when reading a handful of expert-layers immediately after routing), it's the wearout resistance which opens up the possibility of storing KV-cache (including the "linear" KV-cache of recent Qwen, which is not append-only as it was with the pure attention model) and maybe even per-layer activations - though this has the least use given how ephemeral these are.
Is it too late for Intel to bring them back to life?
Yes, their NAND division has been sold, it is now mostly under solidigm. Maybe solidigm could bring it back, but it seems unlikely (given the previous commercial failure).
Nvidia and SK Hynix are bringing HBF to market for $$.
Wouldn't be Intel if they didn't quit halfway through on a good thing.
Still, couldn't one get a RAID 0 card with four drives to saturate a 16x lane? That's already the max one could push through PCIe anyhow.
pmem
This is a pretty cool project! Essentially this is like using Swap memory to extend your RAM, but in a 'smart' way so you don't overload the NVMe unnecessarily.
I do wonder in practice how the 'smarts' pan out, because putting a ton of stress on your NVMe during generation is probably not the best choice for it's longevity.
This is not putting any stress or wear on the NVMe, it's a pure read workload.
Yes, exactly this.
> but in a 'smart' way so you don't overload the NVMe unnecessarily
"overloading NVMe"? What is that about? First time I've heard anything about it.
> because putting a ton of stress on your NVMe during generation
Really shouldn't "stress your NVMe", something is severely wrong if that's happening. I've been hammering my SSDs forever, and while write operations "hurt" the longevity of the flash cells themselves, the controller interface really shouldn't be affected by this at all, unless I'm missing something here.
Hypura reads tensor weights from the GGUF file on NVMe into RAM/GPU memory pools, then compute happens entirely in RAM/GPU.
There is no writing to SSDs on inference with this architecture.
Even if there was a ton of writing, I'm not sure where NVMe even comes in the picture, write durability is about the flash cells on SSDs, nothing to do with the interface, someone correct me if I'm wrong.
People talk about "SSD endurance", but enough parallel I/O on M1/M2 can make the NVMe controller choke, with very weird latncy spikes.
I had assumed heat generation on the controller if it's continuously reading. But maybe it's not actually bad.
Just pop a heatsink on it and call it good.
It will be interesting to compare this to https://news.ycombinator.com/item?id=47476422 and https://news.ycombinator.com/item?id=47490070 . Very similar design except that this is apparently using mmap, which according to the earlier experiment incurs significant overhead.
It was written by an LLM, so... yeah.
Except this isnt using heavily quantised versions of the model thus reducing quality.
Are there any 1T parameter open source models?
Kimi 2.5?
That model is "open weight", not open source. We have no idea what data Moonshot trained on.
Thanks, TIL.
I am curious how the TPS compares vs default OS virtual memory paging
I wonder how many minutes per token on GLM 5.
This is <1 tok/s for the 40GB model.
Come on, "Run" is not the right word. "Crawl" is.
Headlines like that are misleading.
Could still be useful; maybe for overnight async workloads? Tell your agent research xyz at night and wake up to a report.
Assuming 1 token per second and "overnight" being 12 hours, that's 43 200 tokens. I'm not sure what you can meaningfully achieve with that.
Yes, and with virtually zero context, which makes an enormous difference for TTFT on the MoE models.
There needs to be something like this from Ollama. At the moment Ollama has a lot of flaws that prevent it from getting great performance. (My understanding is better GPU/CPU splits, etc). But Ollama is the only way to host an LLM and have it switch out on demand. Sigh.
Ollama has very substandard support for mmap at present, which hurts inference with larger models. There are some recent pull requests in flight that should help address this to at least some extent https://github.com/ollama/ollama/pull/14525 https://github.com/ollama/ollama/pull/14134 https://github.com/ollama/ollama/pull/14864 but progress seems to be stalling. Their support for recent Qwen models seems to also have some bespoke incompatibilities with llama.cpp, which doesn't help matters; it's difficult to test the same model with both.
llama.cpp and llama-swap do this better than Ollama and with far more control.
Don't even need to use llama-swap anymore now that llama-server supports the same functionality.
You do not provide any comparison to llama.cpp with mmap.
You do not explain how any kind of predictor can work for MoE experts.
You do not explain how prediction can even be useful. I can predict the layers used in a dense model (all of them are used in order), but that doesn't help me much. It's still bottlenecked on bandwidth (hint: MoE doesn't change this).
OS paging would be significantly worse here. The kernel's page fault handler is reactive — it doesn't know you're about to read layer 47's FFN weights, so it can't prefetch. You stall on every fault, wait for the 4KB/16KB page to load, then resume. With 80 layers of dense FFN streaming, that's thousands of cold faults per token.
> The kernel's page fault handler is reactive — it doesn't know you're about to read layer 47's FFN weights, so it can't prefetch.
man 2 madvise
That assumes you have significant work to do between fetches (so you can prefetch while using the current data). With LLM decode you don't.
Simon Willison wrote a good post about Dan Woods’ work on “Autoresearching Apple's "LLM in a Flash" to run Qwen 397B locally”.
[0] https://simonwillison.net/2026/Mar/18/llm-in-a-flash/