Is this a derived from flash next? That's the impression I get from reading over everything. Using Strata + Qwen 3.8 Flash Next now and I am beyond impressed. It's basically as capable as 5.6 Luna in my hands so far. If this is an upgrade I can't wait to try it out.
I assume so - Qwen/Qwen3.8-Flash-Next is marked as the "Base Model" on the huggingface page and states it may be subject to "Qwen's separate authorization or attribution requirements"
And I assume the metrics comparing qwen are using a quantized version - from the blog they compare memory use against qwen's "standard 4-bit deployment". Which might explain the improved performance over that base model they show - it might lose less from quantization.
It's interesting. It does seem to be based on QFN, but they've modified the architecture. 10B active versus QFN's 6B, plus also QFN is 180B parameters, not 145B. But it has the same 51B n-gram table.
3.8F/N is amazing. Truly useful, runnable local AI on machines that are not ridiculously expensive. (I’m running it on 2x 3090 and 128gb of ram, on the quantum roughly equivalent to 5-K-M. fast, but you can run it on much less
This looks like it would actually work on my 64mb M1 MBP. That would be fantastic.
> Using Strata + Qwen 3.8 Flash Next now and I am beyond impressed.
The inference engine they use is called edge0-strata because it's a strata fork.
This project looks like someone fine-tuned Qwen3.8 Flash Next, made some tweaks to the model, forked strata to run it, and then put a lot of marketing and branding on top of it
Interesting, if 3rd party testing and benchmarking hold the published values this is very exciting. I am a bit skeptical given the model size, but I hope it turns out to be true!
I'm really impressed with the Blender examples. It's amazing how far "tool use" has come, when it now means using a tool as complicated as Blender to perform work that would take a human days to complete. I'm a little skeptical the actual models are workable, in the sense that a human could come in after and manually refine and apply better textures, etc. Historically, agent-made designs are messy as hell, and by the time you clean them up, you haven't saved any time. But, that's begun to get better; the best models are now very good at making clean shape-based SVGs, for example, that are no messier than what a human would make in Inkscape or Illustrator. But, I'll definitely give it a try.
Do they say how they interact with Blender? Launched within a python interpreter everything in Blender is accessible as a method call. I hope they didn’t do anything more complex than that.
I just gave Fable 5.2 a shot at placing and routing a couple of circuit boards. I'm actually sure it would work fine for simpler/sparser designs but when the component density gets high enough, Fable spends too long struggling in the layer of the layout it's in rather than doing what a human would do, and intuiting that placement, or a component footprint, or different part choice altogether would enable the work to continue.
At first the work looked awesome, and once I started checking it out, I realized I would have to significantly alter the placements and manually draw most of the schematic to get it up to my standards (and hit the performance criteria of the board).
Yeah, I still expect to do a lot of manual work with this kind of thing. By "this kind of thing", I mean any kind of complex visual thing in SVG. I haven't tried any kind of 3D or circuit layout, yet, as it's just not something I do often, but if models could do it, I probably would do more of it...I have a 3D printer, but too slow at 3D modeling to ever want to do anything custom, and I do electronics projects occasionally, but have never had my own boards printed for the same reason. But, I would expect the capability of the models for 3D and circuits to trail SVG, simply because there's certainly less training data and 3D can be wildly more complicated than the average SVG.
It's surprisingly good but you have to have a bit of "theory of mind" to get what and why it can do graphical manipulations. Somehow Claude can draw 3D object wireframes but struggles to know where to draw the leader for a label in a diagram.
Also funny you should say you're too slow at 3D modeling but interested in using LLMs to do it. I'm on the other end. Not speed demon, but generally it's a lot faster to generate the geometry myself than to try to describe it. There's a lot of "how you got there" stuff that matters a ton for parametric modeling. If you want a CAD model that can easily be worked with you have got to be a bit smart about it, and I am not sure I see LLMs having that kind of spatial intelligence + deep physical world context soon. For one thing, it's a lot harder to do RL on.
TerminalBench scores often correspond to real world performance for coding agents that rely heavily on bash.
This score is roughly on par with 5.6 Terra max, which in my direct experience is perfectly capable of driving a cloud agent harness that can do pretty complex knowledge work.
Thy have shipped this as NVFP4 so its Nvidia blackwell only, presumably their inference engine doesn't support anything outside CUDA 13 specified either. Will be interesting to see if this gains some other quants and other inference engines pick it up and we get some standard GGUF files, its a bit of weird release at the moment rather than GGUF file downloads the variants link on HF is leading to a directory of stuff which contains the model spread across a variety of files.
> Sharing matrices substantially reduces the weights that must be stored and moved while keeping the expert count unchanged. Each group of 64 experts shares two large matrices, while each expert retains a small set of parameters that captures its differences.
145B parameters with only peak 29GB memory usage - really impressive, truly something you can actually run on a decent spec MBP
> For Edge0-145B, day-one support covers macOS, Windows, and Linux, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090, and RTX 5090. The release includes one-click deployment tools and hardware guides.
But... there's no guidance on how to run on a mac at all. Weird.
I am not in the material, so perhaps someone else can comment on the quality of the outcome and technique? After skimming a page I am not really sure that the whole model fits inside 29 GB RAM. They seem to use some interesting techniques.
You have roughly a factor 8 in memory usage from 32 bit down to 4 bit precision. That is quite some room already, but not new.
It seems that they radically reduce the model information required per token via the sparse expert approach - 145B->10B parameters - and then use some other approaches I do not understand at that point to make effective use of loaded information.
Most likely a huge part of the magic is keeping the ngram table on disk, as only one row from it is read per token, so there's next to no performance hit.
Qwen 3.8 Flash Next utilizes this, with 40% of its parameters being offloaded to disk, making the effective memory requirements much lower.
It takes things a step further by simply chopping off that half of experts which isn't (or rarely is) activated in coding tasks.
End result is a somewhat competent [0] coding model which runs on a 64GB machine.
[0] it still occasionally does weird stuff like replacing an entire file with just one line from its answer when you ask it a question if you don't explicitly tell it to not change anything or loop itself if the KV Cache is quantised at below Q8_0.
And their model is just derived from Qwen3.8-next-flash which you can see when going to their HF page. Not much details to be found about the company apart from the founder who "was named in MIT Technology Review's Innovators Under 35".
Not much innovation here though it seems. More like a marketing stunt to collect VC.
Qwen3.8-next-flash is impressive though, especially with Strata or Basalt as inference engines. On a 5090 laptop with 64GB RAM one can run NVFP4 quants at 120-140t/s decode and 3000-4000t/s prefill.
On Mac it is much slower and the Kapitsa engine probably the best bet. Despite the claim of 29GB RAM for Edge0, in order to be fast you need to have at least a 128GB Mac.
So did Strata Vs Llama.cpp, yet Strata pulls 2.5 to 3x the performance of the Llama.cpp codebase it was based on.
If they've improved the architecture further by sharing experts, then this might be plausible.
Harness improvements like context pruning, malformed tool call healing also very helpful to boost smaller local models. Doing alot of work in this area lately.
Qwen 27B in Terminal Bench 2.1 73->85 for instance.
For those of us with slow SSDs but a lot of RAM (relative to this model's file size) would this benefit from the extra memory, loading all in, or would I need to set up an old school RAM disk to prevent the swapping?
Their GitHub repo and HuggingFace pages directly state that it's based on Qwen3.8 Flash Next. Their inference engine is called edge0-strata because it's a strata fork.
They fine tuned it and tweaked the architecture.
The real question is if the model performs better on general purpose tasks, or if it was just fine tuned to increase the benchmark scores.
Fair enough. I was just responding to the "just a quantized version of 3.8 next flash" claim. Benchmaxxing via fine-tuning is a very plausible possibility.
I’m admittedly not an expert, I’ve not had the time or energy to try out many of these new American open weight models.
I’ll just say this, we obviously all know they (American open weight models) are playing catch-up to the competition. So if your reaction is “why would I use this over Qwen this, or Deepseek that” you’re really missing the mark.
Yeah, they are all going to be behind on benchmarks for now. Nobody realistically expects some American open weight model to come out of nowhere and beat Qwen/ Deepseek/ et cetera. But there has been a flurry of new releases from this segment, and it’s interesting to me to watch how it develops.
I am once again impressed by the ability of language models to generate huge amounts of relevant-looking text (and diagrams now!) while actually communicating as little as possible.
It's amazing the level of optimizations we are discovering. At this rate, I wouldn't be surprised if the trend starts reversing, ie. instead of releasing bigger and bigger models, companies start releasing smaller, more capable, specialized models for specific applications
Over the past few weeks, I've been putting together a small, 6.5M param model. It is conversational, does tool calling, RAG, mathematical modeling/calculations, basic physical simulations, and multi-step chained reasoning. It's still very much _not_ a frontier model. But it already does more than I could get a qwen 2.5 0.5b param model to do with finetuning just 1 year ago
Most of the improvements come from being able to direct claude to apply different ideas from recent research that keep coming out. AI labs seem to be focused on 1B+ models, so not much comes out about small models, but many of the things being discovered in research with billion+ model, are adaptable to small ones as well
> At this rate, I wouldn't be surprised if the trend starts reversing, ie. instead of releasing bigger and bigger models, companies start releasing smaller, more capable, specialized models for specific applications
Been reading this for years now. Seems both, bigger models optimized for scale AND smaller models optimized for single machines, keep on being released. I don't think it's changing, different markets.
I guess the question is less about what kind of model is viable and more about what kind of application fits any given model. I'm highly skeptical that my employer is willing to pay long-term costs of close to double what I make an hour for token usage, high precision or otherwise. I'm also highly skeptical of the return in some cases. I have to put my codebases in straitjackets tighter than ever to keep them on the rails, and this requires yet more tokens to do so. For premium work, okay, maybe? But it doesn't always excel at that work. For typical work, I don't see it being worth it for the big models.
Based on Qwen3.8-Flash-Next and Strata. They made some interesting tweaks and did some fine tuning on the model, but it's interesting how much marketing speak was layered on their website to hide this fact.
It's not at all unusual for people to fine tune models, tweak them, and fork inference engines. It's highly encouraged and appreciated in the community. What is unusual is seeing someone wrap this all up in extensive branding and publish it online as a company.
As with all fine-tunes, I'd be a little skeptical of the improved benchmark numbers until people actually use it more. It's really easy to take off the shelf models and fine tune them to get slightly higher benchmark scores without really improving them for normal use. HuggingFace is full of fine tunes that advertise amazing results on specific benchmarks but are worse than the base model when you use them for real work.
Has anyone figured out the macOS situation with this? The blog post mentions Mac repeatedly, saying things like "day-one support covers macOS". But the released stuff so far seems to be CUDA only. I see no info on when and if the actual Mac support is coming.
This weird approach is frustrating, because the actual promise sounds very intriguing, so I'd very much like to test this out.
80 comments
[ 0.24 ms ] story [ 76.8 ms ] threadAnd I assume the metrics comparing qwen are using a quantized version - from the blog they compare memory use against qwen's "standard 4-bit deployment". Which might explain the improved performance over that base model they show - it might lose less from quantization.
https://huggingface.co/Edge0/Edge0-145B-NVFP4
https://github.com/Edge0-AI/Edge0/blob/main/licenses/edge0-m...
This looks like it would actually work on my 64mb M1 MBP. That would be fantastic.
The inference engine they use is called edge0-strata because it's a strata fork.
This project looks like someone fine-tuned Qwen3.8 Flash Next, made some tweaks to the model, forked strata to run it, and then put a lot of marketing and branding on top of it
At first the work looked awesome, and once I started checking it out, I realized I would have to significantly alter the placements and manually draw most of the schematic to get it up to my standards (and hit the performance criteria of the board).
Also funny you should say you're too slow at 3D modeling but interested in using LLMs to do it. I'm on the other end. Not speed demon, but generally it's a lot faster to generate the geometry myself than to try to describe it. There's a lot of "how you got there" stuff that matters a ton for parametric modeling. If you want a CAD model that can easily be worked with you have got to be a bit smart about it, and I am not sure I see LLMs having that kind of spatial intelligence + deep physical world context soon. For one thing, it's a lot harder to do RL on.
This score is roughly on par with 5.6 Terra max, which in my direct experience is perfectly capable of driving a cloud agent harness that can do pretty complex knowledge work.
Definitely worth an evaluation!
145B parameters with only peak 29GB memory usage - really impressive, truly something you can actually run on a decent spec MBP
> For Edge0-145B, day-one support covers macOS, Windows, and Linux, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090, and RTX 5090. The release includes one-click deployment tools and hardware guides.
But... there's no guidance on how to run on a mac at all. Weird.
ModelTokenMemoryUsage [Bytes] = Parameters * (BitsPerParameter/8)
For a model with 145B parameters, this means:
You have roughly a factor 8 in memory usage from 32 bit down to 4 bit precision. That is quite some room already, but not new.It seems that they radically reduce the model information required per token via the sparse expert approach - 145B->10B parameters - and then use some other approaches I do not understand at that point to make effective use of loaded information.
Did anyone try the model?
Qwen 3.8 Flash Next utilizes this, with 40% of its parameters being offloaded to disk, making the effective memory requirements much lower.
My daily driver is this:
https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RC...
It takes things a step further by simply chopping off that half of experts which isn't (or rarely is) activated in coding tasks.
End result is a somewhat competent [0] coding model which runs on a 64GB machine.
[0] it still occasionally does weird stuff like replacing an entire file with just one line from its answer when you ask it a question if you don't explicitly tell it to not change anything or loop itself if the KV Cache is quantised at below Q8_0.
They're not hiding that fact once you get to the repo. I mean it's called edge0-strata and the README says it's a fork of strata.
It's weird to see this marketing web page imply that it was all new and novel work that they did, though.
Not much innovation here though it seems. More like a marketing stunt to collect VC.
Qwen3.8-next-flash is impressive though, especially with Strata or Basalt as inference engines. On a 5090 laptop with 64GB RAM one can run NVFP4 quants at 120-140t/s decode and 3000-4000t/s prefill.
On Mac it is much slower and the Kapitsa engine probably the best bet. Despite the claim of 29GB RAM for Edge0, in order to be fast you need to have at least a 128GB Mac.
https://github.com/incoai/splash
Damn, guess I can't run this on my desktop. /s
Their GitHub repo and HuggingFace pages directly state that it's based on Qwen3.8 Flash Next. Their inference engine is called edge0-strata because it's a strata fork.
They fine tuned it and tweaked the architecture.
The real question is if the model performs better on general purpose tasks, or if it was just fine tuned to increase the benchmark scores.
I’ll just say this, we obviously all know they (American open weight models) are playing catch-up to the competition. So if your reaction is “why would I use this over Qwen this, or Deepseek that” you’re really missing the mark.
Yeah, they are all going to be behind on benchmarks for now. Nobody realistically expects some American open weight model to come out of nowhere and beat Qwen/ Deepseek/ et cetera. But there has been a flurry of new releases from this segment, and it’s interesting to me to watch how it develops.
Over the past few weeks, I've been putting together a small, 6.5M param model. It is conversational, does tool calling, RAG, mathematical modeling/calculations, basic physical simulations, and multi-step chained reasoning. It's still very much _not_ a frontier model. But it already does more than I could get a qwen 2.5 0.5b param model to do with finetuning just 1 year ago
Most of the improvements come from being able to direct claude to apply different ideas from recent research that keep coming out. AI labs seem to be focused on 1B+ models, so not much comes out about small models, but many of the things being discovered in research with billion+ model, are adaptable to small ones as well
Been reading this for years now. Seems both, bigger models optimized for scale AND smaller models optimized for single machines, keep on being released. I don't think it's changing, different markets.
Based on Qwen3.8-Flash-Next and Strata. They made some interesting tweaks and did some fine tuning on the model, but it's interesting how much marketing speak was layered on their website to hide this fact.
It's not at all unusual for people to fine tune models, tweak them, and fork inference engines. It's highly encouraged and appreciated in the community. What is unusual is seeing someone wrap this all up in extensive branding and publish it online as a company.
As with all fine-tunes, I'd be a little skeptical of the improved benchmark numbers until people actually use it more. It's really easy to take off the shelf models and fine tune them to get slightly higher benchmark scores without really improving them for normal use. HuggingFace is full of fine tunes that advertise amazing results on specific benchmarks but are worse than the base model when you use them for real work.
This weird approach is frustrating, because the actual promise sounds very intriguing, so I'd very much like to test this out.