Often (most?) of the time "solving" a hard problem is partly taking it apart into separate pieces and phrasing those pieces in various novel ways that allow one to tackle them individually and then at the end it all magically falls into place (in this case after 900 pages).
I think that's really the key here right? It doesn't really "beat" Qwen3.8-27B.. more of like it's on par / a few points ahead on some, behind on others. But contrary to the Qwen (which is dense) it actually does it as an MoE - so less active parameters.
And the fact that they trained it as a looped model in 4-bit native is also kind of impressive to me as well, because it combines recurrence, routing, and narrow-format - all of which adversely impact numeric stability during training.
It beats Qwen on effectively every coding benchmark, which I think are the most relevant ones to HN :)
On the other benchmarks it trades blows or beats Qwen; as Thinky points out, looping helps reasoning but doesn't scale memorization, so tasks requiring broad world knowledge without tools/search won't be helped by looping.
Still super impressive, and probably replaces Qwen 27B in pretty much any workload!
you can also just make qwen looped / carry forward its reasoning from pass to pass. kind of does the same thing and you dont need to change the weights.
Also, for us Apple Silicon users, MoE models fit the hardware better: I have a 64G Mac, and non-dense models like qwen 3.8 next flash, Poolside.ai Laguna XS, and deepseek 4.1 flash 1017 are practical workhorses on hardware that I can afford. I liked inkling-2-mini when I was using it yesterday with pi as a coding harness.
the moe-ness is orthogonal to whether something can loop as least as far as i can tell, or as least as far as i have tried. i have local dense and moe models looped in the way that this trained one does, qwen3.8-27b does it fine and thats good, but the model in question here misses the link that would make the looping kind of neither here nor there, which is you could avoid doing all that fancy loopy machinery altogether and get several orders of magnitude more benefit just by carrying the state all the way through decode between forward passes. i mean you may as well do both i just mean the bigger and lower hanging fruit is on the table here.
when you do the carry forward link which is just a single save-and-write qwen goes from being the guy in memento without any notes to actually pretty brilliant, i havent looked into what it takes to get a clean benchmark on it but i assume itd be as easy as cloning a repo and hitting go, so maybe i'll do that and see what the numbers actually say.
It's anti-ai because people didn't understand code in many (most?) companies already before agents wrote code. Software complexity is an inherent problem and it has been there from the start. Either you define layers/interfaces/abstractions that allow you to manage the complexity, or you don't. Yes, it gets worse if you don't, but blaming the tool for that is an excuse.
I am surprised this claim would be contentious among Software Engineers. However, I will just point to the earliest evidence, since the claim is that this has been systematic from the start:
- Dijkstra's "Go To Statement Considered Harmful" from 1968,
- the "Software Crisis" mentioned in the 1969 NATO conference report
- "Software Crisis" referenced by Knuth on page 263 of ACM Computing Surveys, Vol. 6, No. 4, December 1974
- the destruction of Mariner 1 in 1962 due to "Software Bug"
- Patriot missile failure in 1991 and the Ariane 5 explosion in 1996 (both due to software issues around Floating Point handling)
- the unmanaged complexity crisis of the early times led Robert H. Morris Sr. (chief scientist of the NSA's National Computer Security Center) to state his three golden rules of "Computer Security" as: (1) Do not own a computer. (2) Do not power it on. (3) Do not use it.
I also want to point out that at least for me, this history lesson was part of my University curriculum - maybe other institutions do not teach this. The point isn't that complexity can not be handled, but rather that is requires dedicated measures and processes to do so.
Again, I am honestly surprised to see you disagree with the claim. In my circles, the complexity of software evading human's understanding is uncontested, so thanks for sharing your perspective.
My klaus kot says it's elevation data from a fitness tracker (distance in km, elevation in m, stddev, max-min). It also says there's no way to identify make and model without "seeing the original file" :D
I find physical exercise, not smoking, and talking to people very enjoyable. These things in particular keep me sane and reassured.
When I was an obese smoker who avoided people, I was stressed out by a lot of other things going on. As soon as those stressors were gone, the bad habits also went away without so much effort.
It's worth considering how many of these perceptions of "discomfort" are caused by other aspects of culture, not nature or nationality or era. All my stress was coming from trying way too hard as an insecure younger person.
That doesn't mean you can hit it too. The only case which I remember was well documented was for a turret gun "controlled by satellite" hit the passenger of a standing vehicle from the back of a truck in front of it (nuclear scientist Mohsen Fakhrizadeh, killed in Tehran in 2020 while his wife, sitting immediately beside him, was unharmed). Also those reports were pretty questionable at the time, there were multiple competing narratives, and I'm not sure the satellite gun story was ever independently verified.
reply