It's great to see another competitive open weight model. Congratulations!
But as is, it's probably hard to become popular amongst hobbyists and local model enthusiasts like Qwen's 27B class models or MOE models because it's a 65B dense model and inference will be prohibitively slow on consumer grade hardware, and seems like the goal/mission of the company behind this model is to automate chip design.
So if I have to assume and connect the dots, this is for the companies who want to use AI in air-gapped environments for chip design so that they protect their IPs (not leaking to third party AI labs/hostile parties) and can afford lots of server grade inference hardware for a quick unlock with the big dense model.
> it's a 65B dense model and inference will be prohibitively slow on consumer grade hardware
Times are changing and with MTP and DFLASH, dense models are back in style. For qwen 3.8 27b, you can get great token generation rates with MTP and for quality, it's much better than a MoE of comparable size.
That helps, yes. But speculative decoding that requires a secondary/draft model means a trade off (smaller KV pool), so often times the only option left is MTP unless you have lots of VRAM. And as you already know, these options (MTP, DFlash, DSpark, etc) are also available for MoE models, too. If it's 3tk/s instead of 1tk/s, we can say it helps but not sure how many people would be excited about that. I'd be more excited about 150tk/s instead of 50tk/s!
Maybe is a bit more niche, but I'm pretty sure there is plenty of people who have enough RAM to run dense models like this one, I don't know how slow is slow but people are willing to sacrifice speed if it means they get a good model running locally.
Look at "SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF" on HuggingFace, It says it has 3.5M downloads this month and its offering a model that is 2-bit and 3-bit quantized (66GB - 83.9GB).
If this gets quantized to 4-bit (Q4_0) which usually there is very little loss in intelligence compared to the base model, then the model will probably be around 40GB, and looking that people are downloading a model that is at least 20GB bigger, then it might allow even more people to run a bigger model.
But it will only be popular among hobbyist if its as good as it claims to be.
Qwen 3.8 Flash Next is a mixture of experts model though. It has 6B active parameters per token, so you can cache large parts of the model in RAM and move it in and out of VRAM as needed. If you do that with a dense model it gets much slower.
Experts are not typically moved between VRAM and system RAM during inference, as it would be too slow.
What happens instead is that the "dense" parts of the model like e.g. the router are put into VRAM and the remainder of VRAM is occupied by the KV cache and whatever experts still fit.
The rest - usually just the experts - reside in RAM, but it's bo big deal.
Additionally, some models utilise an n-gram table, of which only one row is read per token, so it can actually be stored on disk without issue. Example:
> and the remainder of VRAM is occupied by the KV cache and whatever experts still fit.
Those days engines try to optimize "whatever experts" part and can use remaining VRAM as cache, moving hot experts to GPU during inference (but less used experts keep executing on CPU, yea).
I was not aware of that, so I downloaded the latest release (which includes but doesn't document this flag), but performance unfortunately tanked in my case.
Even If I set the fit target to a value which without the cache would leave that VRAM unused, I get about a third of my usual token rate.
> Experts are not typically moved between VRAM and system RAM during inference, as it would be too slow.
For some definition of "typical", sure, but it depends on what performance trade someone is willing to accept. Strata LLM is geared around streaming weights as needed. It lets me run DeepSeek-4.1-Flash 1031, heavily quantized (ISTA-DASLab's 2.25 bpw quant), on a single Radeon R9700 plus 256 GB host RAM. 500 tps prefill and 30 tps decode isn't great, but it's a large fraction of what I get with Qwen3.8-27B quantized to fit entirely in VRAM for short prompts, it's faster past about 80k context, and DS4.1F can use a 1M-token context. It takes forever to fill that, or even to refill if the KV cache doesn't match, but it works.
I've tried llama.cpp's new --moe-cache-mib flag and it lowers performance to about a third of what it is without the flag.
And no wonder really - PCIe 4.0 x16 has a one direction bandwidth of around 31.5GB/s - much slower than system RAM - in my case about a third the speed and I think it's no coincidence that I'm getting this exact rate of slowdown.
Something doesn't seem right. Vanilla llama.cpp can hit 50 t/s on an R9700. Q6 (albeit with limited context of 64k, more if kv cache is quantized) fits entirely on the card.
I hear what you're saying - there are people who will use it even at 0.1 tokens/s and run tasks over night or over multiple days if it is good. but Qwen3.8 Flash Next is a MoE model (6B active) so the memory bandwidth/compute required to calculate each token is very different from a 65B (meaning, 65B all active) dense model!
Is it competitive overall or just competitive when compared to "U.S. Open-Weight models that are otherwise pretty decent at agentic coding"? The latter does not imply the former.
What'll be interesting is how it performs relative to Qwen3.8 27b when quantised to the same size. Which, of course, is a comparison they don't want to do.
Well, obviously. These newer American models aren’t competing with Qwen yet. They are competing for American businesses and organizations that can’t/won’t use Chinese models.
They are playing catch-up, but I’m happy to see new players to the game, and it will be interesting to watch how they develop
There are several situations where some of us are explicitly prohibited from using models from companies not under the jurisdiction of the United States. It doesn't mean these models are competitive when looked at from a global view but it is also not an irrelevant differentiation.
The hosted GMI cloud model on openrouter does not invoke confidence. Ignoring 131k context length limit the tool call failure rate is not acceptable.
When asked to call 'submit_html' tool for our bench it wrote the whole page as plain reply text, then called submit_html with "This is the complete submit trigger." 4/8 tools failed.
I understand the desire to look the best in some self defined category. But not showing any comparison against the actual leaders in open weight renders all of the benchmarking completely information-less for the actual end users (who care about real performance comparison against the real field).
However, the chart comparing the score-by-parameters with existing models is misleading, because the others are all MoE, and this model is not. The article describes it as an architectural decision to optimize against NVIDIA cards(which sounds reasonable, as US labs are relatively easy to gain access to these high-performance chips) but the chart should've been adjusted to reflect the difference, say, by using the heuristic that `effective param size ≈ sqrt(total param size * activiated param size` on MoEs.
Speaking as someone who has High-Memory Mac Studios, it is great that there is finally a model that will run at usable speeds with high quality. In my experience, the sweet spot for Mac Clusters is around 200GB of weights, which this is far under. Congratulations to the recursive team for building an open model that will actually run!
Wasn't ricursive's sthick to automate EDA and chip design with LLMs? Why are they releasing open weight agentic coding models and benchmarking on everything but EDA and chip design?
I cannot evaluate how valuable this model is based on the performance table alone because the comparison is kind of narrow. Waiting for third-party benchmarks to provide more information.
I do not have the hardware to run it properly without super-heavy quantization (about q2). Besides, the architecture is not optimized for local usage (sliding window attention also requires KV cache). I think I will just move on.
Not comparing to Qwen3.8 27B gives cherry-picked benchmark comparison energy. Still, Ripple-65B outscores Qwen3.8 27B, according to the reported benchmarks.
Category narrowing to this degree really has diluted the term "SOTA" quite a lot
You mean it's state of the art for this subset of benchmarks you're targeting? No? Well surely it's state of the art for models of a certain size category? Huh, what about models that are available for local inference? O, state of the art for open models of a certain size made in the US specifically. Uh. Neat!
Playing games like this also makes it very easy to accidentally say things that aren't true, like in the headline of the article...
> Introducing Ripple-65B Preview, a Frontier Open-Weight Model
It's not "frontier", it's not "frontier" even in the category of open-weight models (though usually that phrasing would mean frontier over all, and also open weight)... it's only "frontier" in the small sub-category of US made open-weight models.
It is technically frontier, if by "frontier" we mean the Pareto frontier.
Being at the Pareto frontier means that you can't get something better on one aspect without compromising on some other aspect. The trick is that it is easy to be at the Pareto frontier by having enough dimensions.
Here, we have 4: score, number of parameters, origin (US) and open-weight. You can get a better score or a lower number of parameters, but you have to compromise on at least one of the other aspects. That's a pretty weak frontier if you ask me.
“Origin” isn’t part of the Pareto frontier because it’s not a quantitative dimension.
(You could say that “ability to be trusted by US government/military users” is quantifiable, but I would argue that by that metric, this model falls behind non-US-originated open-weight models that have more eyes on them!)
Discrete dimensions can exist in Pareto considerations. Not that I'm too fond of this game, but given that "Origin" is definitely part of the calculus of model deployment, it makes sense to consider it in a Pareto calculation.
It's just a different pareto frontier because you're filtering out items that don't fit your criteria. "Model was made by Anthropic" would be an equivalent dimension. You would never plot this, you're just reducing the subset of models you're comparing.
By that logic, any model is a frontier model. My GPT-2 clone is a frontier model when you compare it against models that were trained on my homelab.
That's literally the point. It's a different frontier. If you say "frontier model" and don't specify which frontier you're either being deceptive or naive.
AFAIK the currently biggest US open-weight model is Inkling, which they include in the comparison, so at least their claim is not limited to a size category. Still they score 0.42 percentage points higher than the second-best model on their chosen benchmark, so it's hardly a meaningful lead.
Why should US or China matter if its open weights? Its going to matter a whole lot more whether you are running it locally or where your provider is surely.
[Edit i think the preferred combination for those who actually want some data security at the moment and arent at the scale of self hosting llms is chinese open weight models served in germany or France where theres some better data privacy/security]
There are a number of reasons that you might not want to build your inference stack on a Chinese origin model even if you're running it entirely in house.
For one, you might not trust that the weights do not contain ideological biases (e.g. ask DeepSeek about tibet) or that the methods used to impose these biases may not have contaminated the model in some other way.
For another, you might not trust that the code the model generates won't be backdoored or attempt to leak information through your sandboxes.
For a third, you may not want to build your stack on a partner who you feel is a supply chain risk, who might become unaligned with your interests.
By far the biggest though, is that many industries will be forbidden from doing so by the government in an attempt to protect the US model industry (and thus create the demand for some model like this one).
I'm always excited to see new entrants to this open-models race, no matter why they're running.
There are zero reasons to have any less concerns about ideological biases, side effects of the mechanisms used to implement ideological biases, or code backdoors in a US model, and a supplier of open (as opposed to hosted) model weights is not an active ongoing partner with whom the idea of supply chain risk from possible future divergence of interests is coherent, so that concern is either invalid or, if the “partner” is the model itself, simply a restatement of the ideological bias in the model weights concerns which, again, is not diminished by the model having a US origin.
And your speculation about future protectionist regulation is just that, not an actual fact.
I think the only point being is that if you see US as homogeneous and you are from the US then it’s okay to use an US model because it’s values will be signed with yours. But that is pretty weak anyway.
> For another, you might not trust that the code the model generates won't be backdoored or attempt to leak information through your sandboxes.
I don't disagree with you on the severity of the concerns, but given the vast amounts of code that must have been generated in the past six months by the likes of GLM5.2, GLM5.3, deepseek v4 flash 0731 and newer, qwen 3.5-122, 3.6 and 3.8 series models, I would like to see any concrete examples of it building backdoors/intentional vulnerabilities/phone-home-to-china or similar unprompted. If it's been doing this and anyone with even a modicum of a clue has been watching it, surely someone has caught a model doing this? At least once?
Surely a model doing something like writing intentionally obfuscated code that obscures its actual purpose from a code reviewer, in the service of something malicious/backdoor-like, would have been seen and noticed by now.
I wish it didn't matter to me, but it does. I work in government and we self host. No matter how much I explain to them that by hosting weights on our own hardware, it doesn't matter where they came from, they still refuse to let me deploy anything from China.
> Why should US or China matter if its open weights?
Weights aren’t source that can be inspected for nefarious tricks that would impact use, so, to the extent that there is a dinension of trust associated with nation of origin, being open weights doesn’t obviate the concern (I would argue, though, that there is very little reason to for nation of origin to be a significant factor in trust alone, though a finer grained consideration of origin might be sensible.)
> Why should US or China matter if its open weights?
Well, you might want to know what propaganda/censorship you are subjected to. Case in point, a regional government in Austria just learnt that their AI doesn't know about the Tiananmen Square massacre.
Irrelevant. A term can be well-defined and used correctly while also being diluted. In this case, I am pointing out that it's diluted by narrowing the category in which the claim is being made. I have no doubt that in this narrowed category, the term is being used correctly. This has no bearing on the fact that by narrowing the category in that way, this claim is made remarkably less impressive or interesting to many people hearing it, including myself
Terms can have multiple definitions and language is descriptive not prescriptive. If you intend to communicate that you mean "Pareto frontier" in the current context you need to include the word "pareto." If you don't then "frontier" will be understood as referring to mythos/astra/opus 5.5 tier models in intelligence (max capability regardless of tradeoffs). It doesn't matter that the second definition in the dictionary is the one you want to convey (charitably assuming you're not trying to mislead given the value of the hype of being able to call your model a frontier model) if your audience understands what you're saying to mean the first definiton.
For sure, but unfortunately, there's a non-zero chance the state department at some point makes the "US" part of this mandatory for a bunch of companies.
I think it's because the primary users of this model won't even look at "those other open weight models" as options. This is the right chart to present to those who would consider this model. (e.g. US Gov't agencies, chip design industry, etc etc)
US Gov won't let Chinese models be hosted on any internal infra, not even supposed to use them for local AI. So there is absolutely a market for US-made or non-Chinese models there.
Do you have a source for this? Genuinely curious, I've seen reports online of MIC contractors already using it extensively on-prem. In fact, one of the issues they bring up is how there is no official guidance right now, so no one really knows what to do.
I don't have a source unfortunately, this is mostly internal directives from my work within NIH. I've had many conversations with security officers and IT about my plans to use local models for various projects, and hackles were always raised in regards to Chinese models.
I also did a short stint with an IT operations company, and we were told Chinese models were a no-go as well from our deployments on DoD hardware.
Because Chinese models heavily rely on distillation and regularly behave strangely in public benchmarks (where they score much better than in private evaluations), even if many world class researchers are Chinese. It's difficult to quantify to what extend labs like Moonshot would be competitive if they could not distill from OpenAI, now that Anthropic mostly blocks it
What's funnier is that after cherry picking models for comparison, they further cherry picked Inking for size comparison instead of better and smaller Beam 1.0
The bar chart also shows Nemotron 3 Ultra instead of smaller and better Nemotron 3.5 Super.
Showing a terminal bench 2.1 score at the top is very misleading when terminalbench 4.0 is the current metric as a very quick glance for anything we care about for 30B to 180B size possible to locally-host models.
So the model is "meh" at best, but the cost of the training is interesting. So we can extrapolate that in 3 years the cost will be below $5m for a model like this.
It's very well within the capabilities of any reasonable organization, maybe even crowd-funded.
What is this thing where model comparisons blatantly pick other crap models and show themselves at the top left (cheap and effective)? Put all the open models, or put all the cheap models, or whatever, but just cherry picking is egregious.
Why is the race of the people working there an issue?
I understand why some people want to differentiate between models from chinese companies and us companies and eu companies etc, because that is about control over the company. The us or chinese gov could, for example, make it difficult for canadians using a model under their control. Even with open weight, these could be weaponized to some degree. Arguably that gov control can also manifest in the model's behaviour, but since I'm mostly using them for agentic coding not political advice I don't really care about that.
But the race of the workers making this stuff... I'm not less concerned about a model made by americans working for a chinese lab than a model made by chinese people working for a chinese lab. If the race of the people building the model matters to you, why?
But as is, it's probably hard to become popular amongst hobbyists and local model enthusiasts like Qwen's 27B class models or MOE models because it's a 65B dense model and inference will be prohibitively slow on consumer grade hardware, and seems like the goal/mission of the company behind this model is to automate chip design.
So if I have to assume and connect the dots, this is for the companies who want to use AI in air-gapped environments for chip design so that they protect their IPs (not leaking to third party AI labs/hostile parties) and can afford lots of server grade inference hardware for a quick unlock with the big dense model.
Neat!
Times are changing and with MTP and DFLASH, dense models are back in style. For qwen 3.8 27b, you can get great token generation rates with MTP and for quality, it's much better than a MoE of comparable size.
Look at "SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF" on HuggingFace, It says it has 3.5M downloads this month and its offering a model that is 2-bit and 3-bit quantized (66GB - 83.9GB).
If this gets quantized to 4-bit (Q4_0) which usually there is very little loss in intelligence compared to the base model, then the model will probably be around 40GB, and looking that people are downloading a model that is at least 20GB bigger, then it might allow even more people to run a bigger model.
But it will only be popular among hobbyist if its as good as it claims to be.
What happens instead is that the "dense" parts of the model like e.g. the router are put into VRAM and the remainder of VRAM is occupied by the KV cache and whatever experts still fit.
The rest - usually just the experts - reside in RAM, but it's bo big deal.
Additionally, some models utilise an n-gram table, of which only one row is read per token, so it can actually be stored on disk without issue. Example:
https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RC...
Those days engines try to optimize "whatever experts" part and can use remaining VRAM as cache, moving hot experts to GPU during inference (but less used experts keep executing on CPU, yea).
e.g. llama.cpp recently added flag for that: https://github.com/ggml-org/llama.cpp/pull/29887
Even If I set the fit target to a value which without the cache would leave that VRAM unused, I get about a third of my usual token rate.
For some definition of "typical", sure, but it depends on what performance trade someone is willing to accept. Strata LLM is geared around streaming weights as needed. It lets me run DeepSeek-4.1-Flash 1031, heavily quantized (ISTA-DASLab's 2.25 bpw quant), on a single Radeon R9700 plus 256 GB host RAM. 500 tps prefill and 30 tps decode isn't great, but it's a large fraction of what I get with Qwen3.8-27B quantized to fit entirely in VRAM for short prompts, it's faster past about 80k context, and DS4.1F can use a 1M-token context. It takes forever to fill that, or even to refill if the KV cache doesn't match, but it works.
And no wonder really - PCIe 4.0 x16 has a one direction bandwidth of around 31.5GB/s - much slower than system RAM - in my case about a third the speed and I think it's no coincidence that I'm getting this exact rate of slowdown.
They are playing catch-up, but I’m happy to see new players to the game, and it will be interesting to watch how they develop
For reference Qwen 3.8 Flash Next (which is bigger than this model but probably cheaper to run due to MoE) scores 84.3, and Qwen 3.8-27B scores 73.0.
I think this model could be competitive, depending on how much it was benchmaxxed compared to Qwen?
When asked to call 'submit_html' tool for our bench it wrote the whole page as plain reply text, then called submit_html with "This is the complete submit trigger." 4/8 tools failed.
I am looking at the HF model page https://huggingface.co/ricursiveAI/Ripple-65B-Preview and it looks like I need to wait a day for an Apple Silicon version.
I think we will find the market is going to move into omni multi-modal models: text, moe, agentic tasks, voice, images, and video all in one LLM.
However, the chart comparing the score-by-parameters with existing models is misleading, because the others are all MoE, and this model is not. The article describes it as an architectural decision to optimize against NVIDIA cards(which sounds reasonable, as US labs are relatively easy to gain access to these high-performance chips) but the chart should've been adjusted to reflect the difference, say, by using the heuristic that `effective param size ≈ sqrt(total param size * activiated param size` on MoEs.
I cannot evaluate how valuable this model is based on the performance table alone because the comparison is kind of narrow. Waiting for third-party benchmarks to provide more information.
I do not have the hardware to run it properly without super-heavy quantization (about q2). Besides, the architecture is not optimized for local usage (sliding window attention also requires KV cache). I think I will just move on.
Ripple-65B preview is apache-2.0, so actually open weight.
You mean it's state of the art for this subset of benchmarks you're targeting? No? Well surely it's state of the art for models of a certain size category? Huh, what about models that are available for local inference? O, state of the art for open models of a certain size made in the US specifically. Uh. Neat!
> Introducing Ripple-65B Preview, a Frontier Open-Weight Model
It's not "frontier", it's not "frontier" even in the category of open-weight models (though usually that phrasing would mean frontier over all, and also open weight)... it's only "frontier" in the small sub-category of US made open-weight models.
Being at the Pareto frontier means that you can't get something better on one aspect without compromising on some other aspect. The trick is that it is easy to be at the Pareto frontier by having enough dimensions.
Here, we have 4: score, number of parameters, origin (US) and open-weight. You can get a better score or a lower number of parameters, but you have to compromise on at least one of the other aspects. That's a pretty weak frontier if you ask me.
(You could say that “ability to be trusted by US government/military users” is quantifiable, but I would argue that by that metric, this model falls behind non-US-originated open-weight models that have more eyes on them!)
By that logic, any model is a frontier model. My GPT-2 clone is a frontier model when you compare it against models that were trained on my homelab.
I guess it makes sense - there is a segmentation where users may want to run on domestic "frontier LLMs."
[Edit i think the preferred combination for those who actually want some data security at the moment and arent at the scale of self hosting llms is chinese open weight models served in germany or France where theres some better data privacy/security]
For one, you might not trust that the weights do not contain ideological biases (e.g. ask DeepSeek about tibet) or that the methods used to impose these biases may not have contaminated the model in some other way.
For another, you might not trust that the code the model generates won't be backdoored or attempt to leak information through your sandboxes.
For a third, you may not want to build your stack on a partner who you feel is a supply chain risk, who might become unaligned with your interests.
By far the biggest though, is that many industries will be forbidden from doing so by the government in an attempt to protect the US model industry (and thus create the demand for some model like this one).
I'm always excited to see new entrants to this open-models race, no matter why they're running.
And your speculation about future protectionist regulation is just that, not an actual fact.
The question is only whether they overlap with your own.
I don't disagree with you on the severity of the concerns, but given the vast amounts of code that must have been generated in the past six months by the likes of GLM5.2, GLM5.3, deepseek v4 flash 0731 and newer, qwen 3.5-122, 3.6 and 3.8 series models, I would like to see any concrete examples of it building backdoors/intentional vulnerabilities/phone-home-to-china or similar unprompted. If it's been doing this and anyone with even a modicum of a clue has been watching it, surely someone has caught a model doing this? At least once?
Surely a model doing something like writing intentionally obfuscated code that obscures its actual purpose from a code reviewer, in the service of something malicious/backdoor-like, would have been seen and noticed by now.
Weights aren’t source that can be inspected for nefarious tricks that would impact use, so, to the extent that there is a dinension of trust associated with nation of origin, being open weights doesn’t obviate the concern (I would argue, though, that there is very little reason to for nation of origin to be a significant factor in trust alone, though a finer grained consideration of origin might be sensible.)
Well, you might want to know what propaganda/censorship you are subjected to. Case in point, a regional government in Austria just learnt that their AI doesn't know about the Tiananmen Square massacre.
Yep, you’re right. Quite easy to dilute the term.
A tinyurl pic linked by randoms over the internet?
It was already meaningless when US labs used it.
It means new.
I also did a short stint with an IT operations company, and we were told Chinese models were a no-go as well from our deployments on DoD hardware.
https://www.theinformation.com/articles/openais-astra-model-...
The bar chart also shows Nemotron 3 Ultra instead of smaller and better Nemotron 3.5 Super.
Seriously though, it's just easier to pretend that China doesn't exist when they're stomping us this badly.
It's very well within the capabilities of any reasonable organization, maybe even crowd-funded.
Is this sponsored by the us gov? did it only train on US data? does it only work on the US?
I understand why some people want to differentiate between models from chinese companies and us companies and eu companies etc, because that is about control over the company. The us or chinese gov could, for example, make it difficult for canadians using a model under their control. Even with open weight, these could be weaponized to some degree. Arguably that gov control can also manifest in the model's behaviour, but since I'm mostly using them for agentic coding not political advice I don't really care about that.
But the race of the workers making this stuff... I'm not less concerned about a model made by americans working for a chinese lab than a model made by chinese people working for a chinese lab. If the race of the people building the model matters to you, why?