Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s | Hacker News Reader

Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s

github.com

240 pointscarloslfu118 comments

I built slotstream, a way to run Qwen3.8-Flash-Next 4-bit on a low-memory mac starting from 16GB, a 125B parameter model that would need 100GB+ memory/RAM, thanks to expert-offloading/ssd-streaming. Easy to install/update, and mac-native using MLX and Swift.

It ships with auto-mode, which makes a good tradeoff between memory usage and speed. I'll be implementing and porting the MTP module for speculative decoding next

Open articleView on HN

> Hugging Face is the bottleneck, not your link.

README could clearly make use of a cleanup, seems to be more like a session log dump now than a good introduction to the project for a new user. Maybe try something like "Remove anything from the README.md that wouldn't be helpful to someone who sees this project with zero context, for the first time. Rewrite all paragraphs and sections to be concise and remove all fluff, leave only important details new users must know before using the project".

thanks! I'll do!

I hate this AI style writing because since it doesn’t really understand flow, it’s being inserted in irrelevant places and it is extremely irritating to read.

I feel you! fix incomming

For what it’s worth, this comment was not targeted at you, but rather the model kinda forcing it. I get the sense that Anthropic did not think much of this, but it seems to have gotten worse with recent models and it really comes off as a kind of nails on the chalkboard writing style.

I have to image whatever style of writing this was trained on is a lot more pleasant to read and I feel bad for whoever writes like this now being associated as bad AI writing.

most of this stuff isn't really possible at this point _without_ models, so pissing and moaning about people publishing work and not humanizing is really silly.

I used similar prompts before. Now I simply say to "remove historical cruft" and results are good enough. It's the model itself that first used this wording, I found it concise.

I don’t want to be a cranky codger, but I dont get why 5 minutes of work cleaning up the README can’t be done before posting to HN.

Honestly? Me neither, I'd be embarrassed if I got caught not even looking at my own work before shipping it publicly.

However, feels like the battle is loosing, and now it's just about softening the landing instead of being vehemently against something the vast majority seems to be engaging in. I refuse to participate myself, but at least hopefully I could help steer some of them in a better (more concise) direction.

And it can even be LLM work. Just write a good requirements doc for document creation, including links to good readmes, maybe a style guide for your personal style.. and you can have quality docs with near zero tax.

Sure, hand editing each one is even better, but you can get 80% of the way there with zero ongoing investment.

> Run Qwen3.8-Flash-Next on a Mac that can't hold it

This is the first line of the README. I can't believe people are becoming ok with this, and I'm 100% on the AI train.

I have a 48GB M5. I don't need to run larger models. I want more context. I've managed to set the context window at 71,680 using Qwen3.8-27B-oQ4e-fp16-mtp. But I want more. Is anybody, with similar specs, able to set their context window higher?

yes, with qwen3.8-27b-4bit run via rapid-mlx i can get to about 200k

Nice.. and how many toks/sec?

Around 20

You should try Glimmer MTP. Qwen3.8 27B seems to have weird memory and caching issues on oMLX

We are running 35b-A3b with 264k context (the model's default max) using vllm and the "frog" jinja templates: https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates and had good luck. We are mostly running agentic workloads though, rather than coding. 27b has a slightly higher agentic job completion rate (95% vs 92%) but the 3% trade off is worth it because the A3B is sooooo much faster, and we reprocess the other jobs with a different model. Don't sleep on the froggeric templates.

Qwen: Looking at you for a new ~35B MoE! Please and thank you

I am running 3.8 27b at q6 quant with 160k context on a 32gb video card (arc b70 pro) - I quantized the kv cache at q8 - that is the only trick really - works great.

How are you finding it when it gets that high anyway? I’ve got a 64gb Mac so the context _can_ go higher. But I find around 70-80k it goes a bit weird and starts arguing with itself and doesn’t seem to know who it is vs me anymore.

It's hard to believe 16GB unified memory will give you 5 tok/sec unless you are ignoring the thermal warnings. I am running Qwen3.6-35B-A3B on my 16GB M3 and get 7-8 tokens/sec with all the optimizations while keeping the peak memory and thermal warnings at check. https://github.com/deepanwadhwa/samosa-chat

interesting! Yes, thermal is important. Pretty cool project man! Starred and checking it out!

Now I'm feeling pretty good about getting 10-11 tokens/sec running Qwopus 3.6-35B-A3B Q6_K on an old Mac Pro 2013 (trashcan) with 128GB RAM (DDR3), 12 core Xeon, dual D700s. Arch Linux and llama.cpp.

haha, good for you.

Anything smaller than a 16” runs into serious thermal problems; even an identically equipped 14” just can’t dissipate enough heat.

The laptops definitely can't hang but the minis don't really care. I threw mine down in the basement just to put the heat somewhere else, can tell when the dehumidifer next to it is on because it's a few C lower but that has no impact on performance. I don't think it's ever seen anything north of 70

I'm hoping to see progress in this space.

Folks talking about how 32G is not enough for local use, but then there's been work like this to empower it.

My hope is that the new 32G M6 will be "useful" locally, possibly because of work like this.

yes! I'm bullish on this. there is a lot of work to do. I've been experimenting with pruning, distillation, and retraining too. I'm sure your 32gb m6 will run a badass local model!

Yes, but also 12 tok/s versus Claude is so far from comparable. I know that it’s not exactly 1:1, but it’s a long way from an easy trade-off, especially considering hardware prices for high levels of RAM.

32GB is simply too tight; you need 8 minimum for the OS and you need about 4-8 more for the LLM you’re visiting and kv cache.

Not a mac/UMA discussion point, but is it time to add additional, installable, DDR5 to GPUs? I can see this as a win/loose. PCIe 5x16 is close to maxing out the bandwidth available from high end dual channel DDR5 now, but not quite. I'm not a hardware person but I suspect putting it on the card could lead to significant performance improvements over using system ram so allowing systems like this, where MOE weights are shed, to get even higher performance than just adding that DDR5 to the system. Bigger models become closer to reality and it provides more of a pathway for developing technologies that take advantage of it. Of course the loose side is that you just put a lot of specialized ram on a card instead of into the system where it could be used for other things. I could see a place for a 16GB card with 64GB(or more) of DDR5 especially if we start seeing MOE and similar technologies really start being designed for this concept.

Probably not with DIMM modules, as longer traces mean higher latency (speed of light is ~30 cm in 1ns). GDDR typically uses larger buses (more wires) for higher bandwidth, even more so for HBM, so DIMM would be hard. Maybe CAMM would be up to the task?

It certainly seems feasible from an engineering perspective (though it does make cooling harder), at least for mid-range, not H100-class HW, but it prevents market segmentation, so EOMs may not be too interested (as long as no competitor does it).

"Disk is the gate that bites first"

AI;DR

The never ending gate bites.

How I have come to detest certain phrases.

Load bearing gate bites.

...the seam.

Interesting! You're absolutely right, I'll check

Ha! AI;DR is a great phrase. Had not seen that before

I love these efforts to get proper models running on lower cost hardware and I think this is where the next real breakthrough will come from. The more efficient this sort of thing can be done the bigger the chance to democratize this tech, 'good enough' is what you need and as long 'top of the line' gives a competitive edge even if it is at a cost there is a substantial risk of the door closing on general computing at some point in the near future. Keep in mind that there is no guarantee that the pendulum has to swing back, it can swing one way and get stuck, and then you're going to have to beg for crumbs from the haves.

I think there's a very good chance that history will rhyme a bit.

DOS/Windows and PC clones were by no means the best available, but they were cheap, ubiquitous, and versatile compared to alternatives that were either much better at one task but more expensive or better at everything but wildly expensive. They were "good enough" and represented a solid improvement over what many existing computer users had as well as a good entry point for new users. As such they spread like wildfire and became the standard while the expensive alternatives either became hardcore niche or vanished.

SUN Apollo SGI

Though to be fair it was Linux more than Windows that killed them. Dos and Windows were competition for DEC and - ironically - IBM.

I am trying to make it easier to use LLMs on older, cheaper, smaller GPUs. I'm taking a similar approach (move MoE expert weights to disk, avoid wasting VRAM on these). My goal is also to run models that do not fit. My work also suffers from AI documentation issues. Where my approach differs is that instead of running an LLM that doesn't fit slowly, run many agents in parallel sharing the streams of MoE experts weights, to increase throughput. I envision a team of AI agents sharing a pretty-good-at-coding LLM that does not fit to collaborate on a set of related features, being developed in parallel.

My work is showing promising results (if you can get past the way the AI tries to describe what I am doing). https://sw-ml-study.github.io/emufpga/index.html

I am doing this work initially on a 6-Xeon-cores Linux workstation with an RTX5060-16G to run MoE models larger than that. Then I will be moving this to a server with a lot more cores (Dual 32-cores) and a mix of SAS HD and SSD drives, using older GPUs.

Ultimately, I hope to build some FPGA/MCU "accelerators" that process the expert weights on systems with not enough CPU cores to offload the experts. If I can enable large capable models to run on older hardware, keeping the limited GPU VRAM for context and things that must be in VRAM, I can get useful work out of my old refurbished systems without paying today's RAM and VRAM/GPU prices.

As someone who is just looking at the theoretical benchmarks of each of these models I'm curious if anyone could share what are the problems (maybe around code) that flash-next was able to solve which 27b was not able to

This is the best I could find: https://huggingface.co/Qwen/Qwen3.8-Flash-Next?utm_source=ch...

About the specifics, I have only anecdotal evidence, but I guess this info can be found somewhere

For a local non-coding agent, instruction following and tool use are the most important gains

Brilliant! Have you had any success integrating it into a MacOS Swift app. I'd love to see it in action before I consider adding it to a future build. My biggest issue is getting these opensource models to use tools well enough for production.

This is next! in the works rn.

I tried this in my mac mini m2 16GB, unfortunately I have to use an usb disk for the model weights, and I’m getting 0.5 tok/s. Still, being able to run (heh maybe crawl is more accurate) a 100B model on this computer AT ALL is pretty cool.

I see disk maxing out at 400 MB/s, this disk should be able to hit 1GB/s (it hits that eg when verifying the check sum of the weights), so there might be some optimization to be done there (I’m guessing it’s because the weights access is not pure sequential reads but involves some randomness depending on which expert)

For what it's worth, I've spent some time with Claude to develop a local runner for `llama.cpp`. I run Qwen3.6-35B-A3B-MTP (fast!) and Qwen3.8-27B (20 tokens/s).

This was definitely worth the effort. Benchmarking and testing various approaches and various options really paid off. For example, one thing that surprised me was that MTP made things slower, not faster for Qwen3.8-27B.

I use a 64GB MacBook Pro (M4 Max).

I find mtp=3 does well with that model, only at 4 it becomes unprofitable.

Check your quants, its worth having the mtp layer be a bigger quant if it leads to 2x throughput from more accepted tokens.

I’m not an expert, but my understanding is that MTPs are smaller LLMs fine-tuned to "mimic" / predict a specific model’s response. It’s possible that the MTP you’re using isn’t trained well enough on Qwen 3.8. What accept rate are you getting?

For qwen, it's an additional transformer layer at the very back, it ships as part of the model.

Acceptance rate is good, but MTP doesn't help in my case because of my machine's memory bandwidth constraints (M4 Max). Turn out it's better to turn MTP off.

how much energy does it consume?

Good one! I haven't measured this. I'll include it!

Also I don't know if people like this... but could this be wrapped around a nice desktop app?

great idea!! a native app would be awesome

It seems we could use a new kind of memory that streams the weight data in, like GDDR in reverse.

High Bandwidth Flash? https://www.sandisk.com/company/newsroom/blogs/2025/scaling-...

interesting!

Optane? (Too soon?)

Optane was targeting the latency gap, while HBF targets bandwidth. Given how LLMs work, HBF is perfect for offloading.

yes! I guess future hardware designs will have something like that!

I swear the models are named by the beatbox aliens from the post office in MiB

How usable is 12 tok/s?

Unpleasant for interactive agentic work. Still useful to leave it to do some work in the background.

32GB dedicated to an N-gram table instead of a draft model is an unusual choice for speculative decoding — what made it win over the more common draft-model approach here?

It wasn't either/or, the N-gram table is part of Qwen itself and stays on disk. I’ve now added its 1.5GB MTP draft head too, it gets 86% acceptance and about 1.24× faster decoding on my 48GB Mac.

Got it, thanks for clearing that up. 86% acceptance is solid — does it stay flat over longer generations, or drift with context length?

Would love to see metrics of model performance, comparison with oMLX/OLLAMA/others. Any tooling for local optimization on hardware.

To be more clear, I see you have slotstream doctor --sim-ram N, but extended tooling and optimization for exact local hardware would add value (MTPLX has a nice interface for example).

so many inference project, omlx already supports all of this and has a 1000 people trying to optimize it constantly

Interesting! I'll check it out

Both projects are different in scope. Think of slotstream as optimizing for memory and for this specific model for now, my intention is not to build an inference engine the same as oMLX

I just used it, went through the whole installation (took like 1 hour approx). Long but straightforward. If I put my computer to sleep will it continue?

I started the server (very curiously I was running oLlama in the same prot slotserve uses by default, instead of switching it which I know you can do, I just ditched oLlama, perhaps an insight for you) and built a small html hello world served via Python. The thing pointed me to the localhost link, nice!

As an early user, my advice is to focus on efficiency. The efficiency of the installation but more importantly, the efficiency of running the thing. 8.1GB per slotserve process is a lot! Is that in your control?

Also, I've seen an interest of certain kinds of programmers for open-weight models. "We all know agree that LLMs for coding are very useful but we're giving money to a small set of big, evil corporations. They're Trump donors. I heard it's bad for the environment because it uses water". If it's local and open-weight, this could be marketed this way I think.

Finally, what's the actual, real use case for slotserve?

thanks!

> ditched oLlama"

yeah! this is interesting.

> 8.1GB per slotserve process is a lot! Is that in your control?

yes, it is hard, but I agree the smaller the better. I'll work on that

> If it's local and open-weight, this could be marketed this way I think.

I like this!

> what's the actual, real use case for slotserve?

I'm working rn on an app on top of it that closes the loop and is a fully local AI app, an experiment. I'll publish it as soon as it is usable!

> built a small html hello world served via Python

What did you use as a harness here?

For the harness... just the shell. No client library. Does that answer your question?

thanks! in part, I was wondering how you got the code into files. I guess you copy pasted it inside a file, am I right?

Is this going to destroy my SSD?

By reading it?

I don't know actually. I'll check haha. My best guess is it isn't.

I hope not! this is a new macbook lol!

Reading causes insignificant wear ("read disturb") that likely isn't a problem, but I don't think it's possible to issue pure reads to modern SSDs. The NVMe spec mandates tracking the amount of data read, and this has to be written to the drive. I'd hope the firmware buffers this and writes it at low frequency, but on the other hand, I doubt the firmware was tested in extreme random-read regimes. Unexpected failures from excessive statistics recording could be possible.

no it's reading, not writing

Using macos on low memory regimes will make it use disk-based swap.

For example, a Macbook Neo (so in theory, something with around 4GiB of free RAM lying around) might eat around 900GB of writes a day while not doing much at all, because it's basically on low on RAM and swapping all the time.

A particularly worrying situation considering a dead SSD will render your macbook usable for parts only.

Next help us normies run GLM 5.3 on our potato computers. Wouldn't that be nice!

Colibri did that first for GLM-5.2 https://github.com/JustVugg/colibri

There are already a handful of repos doing essentially exactly this: `mlx-moe-offload`, `streamlx`, `mlx-moe`, `mlx-flash`, and `deepseek-v4-flash-mlx` - i.e. keep the resident parts of an MoE in unified memory and page/stream routed experts from SSD on Apple Silicon.

At this point I'd much rather see people collaborate on one of these implementations, benchmark against them, or upstream the useful bits into MLX/MLX-LM instead of producing yet another near-identical repo.

The local-LLM ecosystem really does not need every implementation idea rediscovered five times and wrapped in a new README. AI-assisted coding makes producing a new repo cheap; maintaining, benchmarking, and integrating one is the actually valuable part.

> every implementation idea rediscovered five times and wrapped in a new README

That's open source since forever, unfortunately.

It's what happens when you don't do market research.

I'm sorry this makes it seem like I didn't do my research. I did a TON. To fix it I'll add a benchmark/comparison table. Also, I wouldn't call it market research since this is not commercial AT ALL.

Hey Carlos, thanks for sharing with the community! Appreciated

thanks!

Does a painter check to make sure that a portrait hasn't been painted? What a dismissive comment.

Half the people on here are using Ollama. No one is doing market research.

I agree with the sentiment, but have you seen those videos in which all men say other men are gay? This feels like the same, so much AI paranoia!

I genuinely want to contribute. And hey! I was doing oss this since 2014 so waay before AI was cool.

Vouched especially since OP might have a perspective on this. And readers may want to look up those other repos and compare for themselves.

Thanks for the feedback! I'll create a section with a benchmark and comparisons. This will hold the project accountable and speed things up imo

Absolutely! Nice.

(The comments under the parent indicate it was improperly flagged/made dead (maybe could happen just from downvoting?) so glad I hit the Vouch.)

AI NIH

Sorry, I don't get "NIH". what's that?

Not Invented Here.

Ah! Yeah, I didn't invent anything (yet!). The goal is to see how far I can take it in terms of speed without consuming that much RAM.

I'm only joking anyway — it's more a comment on the whole AI-accelerated trend of everyone having their own version of a thing.

I do agree that, ultimately, combining your efforts with others working in this whole area is probably really worth it, but I can see how there's an ease of pushing forward on your own these days.

I do not have fast internet so I am not sure when I'll really be able to download the weights but I do have an M1 Max to try this on, so I will at some point!

I see! yes, downloading the weights part is painful. I tried a couple fixes and it is as fast as it can get downloading from HuggingFace. I think the field is heading toward smaller, more capable models soon, so you won't have to wait that long!

I mostly use Muse Glimmer, which runs quite well on my M1 Max as it is 30B but it also reasons very efficiently. I have tried the Qwen 3.8 27B which is also usable if quite slow to reason, but I guess like everyone the idea of the Flash-Next model holds some intrigue, because the 35B-A3B is pretty good.

I see your point. As an oss defender myself, I agree, however, the spirit of this is to see how fast I can make it. I'm sharing this with the community, which I think is aligned with the original oss spirit.

It's an experiment for myself but I am committing to maintain it. I've been an oss person for a loooong time, way before AI was a thing. Think about it as a new, from-scratch take at it, not as a re-reproduction.

Hey carloslfu, kudos from the other side of the internet, don't get down on people nitpicking everything here, experimenting and discovering is part of learning so keep going!, remember this is the place that said dropbox was dumb and could be replaced by a script.

And there are `Mference` and `SwiftLM` too, I think they are doing the same use case.

Why should they do that? For you? You could merge those projects and see if they get traction.

This is one of the aspects of this year that I've been finding very grating and wasteful. Collaboration still happens among people with the ability to do so and the technical skills, but everyone else is taking their own helicopter to the top of the mountain, "putting it out there", and there's just a ton of redundant projects that do the same thing.

It's horrible. Every 20-something working on a load-bearing inference engine on GitHub.

for the record, I'm almost 35

Also for the record, I wasn't trying to take a personal shot at you or your project—nor am I sure of how valid that would be, if that were to have been my intention—it's just a gripe I have in general what I think is somewhat degrading the trust I can have in certain types of projects, especially those that someone shits out, puts on the app store, appears on the surface to visibly look ok, but ultimately has no uniquely valuable contribution or long-term outlook and is just someone's zero to one replication of something that has an api and already exists. It is admittedly cynical, but I now scrutizinize what I pay for more aggressively as a consequence.

I think multiple people working on the same thing is great.

Everyone comes at it from a different point of view, and some approaches work, some don't. And when people do this themselves they learn. Existing projects have their mistakes worked out already.

Maybe one of these people is going to come up with the thing that nobody else thought of because of their experience working the problem from scratch. You may not get that from someone working from an existing project, because existing projects have their approach "baked in."

What all these projects are showing so far is that it's possible to stream from disk, but that the performance isn't ideal. But I'm sure you could take this approach with smaller models and get better performance.

In addition, it's a given that when you work with large data sets performance means organizing the data to take advantage of caches, both disk and cpu. It's not clear how that would work, exactly, given that each run is a not-quite-random walk through the data. The Big Data way is to prebuild all of that as much as possible, which is probably impossible with a big model. But what about a smaller model?

true

Multiple people working on the same thing is great. I'm less convinced about multiple people asking the same LLMs to redo the same project and coming up with a repo with a llm slop readme full of "Disk bites first", it's not this, it's that etc...

I don't think that this will bread any innovations.

Reply on news.ycombinator.com