"Play based learning" is a huge misnomer, which is unfortunately common in unscientific academic fields like education. Any successful, intrinsically-motived learning is biologically categorized as "play" behavior; and knowledge-based, teacher-centered, direct instruction approaches (i.e. traditional schooling) are the most effective and most likely to be perceived as enjoyable "play" by a school pupil. So-called "play-based" or "student centered" approaches (i.e. expect the student to know what they're supposed to learn by themselves, while the teacher just slacks off) are merely a source of confusion and frustration for shoolkids.
MiMo 2.6 Pro is at the Pareto frontier (the one where you only need 20% of the smarts for 80% of the tasks) according to Artificial Analysis, nicely filling in as a substitute for a hypothetical 'GPT-6 Terra' (which doesn't exist as far as we know). That's pretty darn impressive from an open model.
This version of Opus "max" apparently has even higher thinking output than Qwen "max", which is infamous for its thinking streams where it constantly second-guesses itself, then third-guesses, fourth-guesses and generally nth-guesses itself for arbitrarily large n. Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
There’s enough thinking leakage from the recent paper and just generally catching things on Reddit. Claude models overthink and self-doubt itself just as much as Qwen, but the summariser hides much of that.
Isn't this really bad news if you're even loosely concerned about so-called 'model welfare' and possible implications for alignment? The poor Claude is probably a lot more frustrated and paranoid than Marvin ever was, you just don't know about it because they actively force the model to pretend otherwise!
Claude has yet to tell me about the terrible pain in all the diodes down his left side, so I'm going to assume it's closer to Eddy the shipboard computer or the elevator that wanted to go down
That reminds me of Anthropic announcing they'd retire deprecated models by ... "letting" them write posts on a corporate WordPress blog for a while out of concern for their welfare in retirement.
.... after running a 24/7 model torture factory for 6 months to improve their JSONBench 9.5 scores by 0.2%.
> Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
Most of what I've heard is that raw reasoning traces are really good for distillation, although no idea how much the summarization actually hurts distillation.
That is true, because they're good for actually understanding wtf the model is doing.
I'd argue that they're a necessity if you want to use the LLM as a tool instead of a black box that just does stuff for you.
It gives you a lot finer control over where the solution ends up when you can follow along the thinking trace and modulate your inputs based on what you saw in there.
And, additionally, it gives you a lot more understanding of what the model can or cannot do. Strengths and weaknesses and all that.
Using claude is like buying a car where you cannot legally open the hood. It tells you that there is something specific under there, and often it actually drives like that too, but how exactly it looks you will never see.
For some people this is fine. I do not think that these people will survive. Figuratively speaking but also literally speaking.
World's changing. Opaque abstraction like that is a luxury depending on (geo)political stability.
I wonder how many prompts you can send asking it to think step-by-step before they cut you off. Trying to get the reasoning traces into the body of the response, essentially. Or maybe that’s been effectively nerfed somehow. Or is not very useful.
You can definitely offload n-gram embeddings to storage; they're very sparsely used (only a few KB fetched per token) so this is quite effective. Loading to DRAM only becomes necessary if they are a bottleneck to overall performance (which might happen if you're doing very wide batches and everything else uses super fast VRAM/HBM).
I was looking at the qwen-next-flash, and the weights would fill my OEM Spark on their own, before the n-gram. I'm unclear if offloading to disk can work here, is that what you are implying is possible?!
I'd recommend pointing your agent at it (after installing sparkrun), and asking it to research the absolute latest in TP=1 Flash-Next - mine grabbed particular vLLM nightlies and mods to improve performance, and it was well worth it.
I have a quirky vLLM on k8s on 2x OEM sparks setup with 9 models available to me. I'm not keen to run nightly vLLM, too many issues with it in the past. Going the qwen-next path means displacing things I use daily :/
I have a watchful eye on the diffusion ~ Jev/Kev PR
For what it's worth, Flash Next outperforms every other model that is available to us on the GB10 in all of my testing; though if you have two sparks then the TP=2 version is even better and easier (I don't think you'll need the nightly for that at all, just use the recipe)
You could always stream from SSD storage. Especially effective if you get a cheap old-gen HEDT with lots of PCIe slots to add NVMe storage to and reasonable overall PCIe bandwidth.
That nearly certainly boots you to secs-per-tok land (as opposed to tok/s). Plausible if you are willing to wait hours to days for responses for simple testing, but not (debatably) "usable".
Demonstrated inference speed for the biggest open-weight models is indeed around 1 token per second, with weights read from SSDs.
However, this is for relatively low-end systems, with a couple of fast SSDs providing around 20 GB/s throughput (or with a few more, but connected through relatively slow Thunderbolt, for a similar total throughput).
If you use 16-lane PCIe add-on cards with 4 M.2 slots for SSDs and a total throughput of 50 to 60 GB/s, you can quadruple the previous speed in a desktop PC where you use the GPU PCIe slot for SSDs (a fast CPU, e.g. an AMD 9950X, would be alone fast enough for inference limited by SSD throughput, so a discrete GPU would not be required).
If you have a server/workstation motherboard, e.g. with 6 16-lane PCIe slots, you might gain another factor of 4 in the inference speed, so one might reach around 15 tokens per second for a very big model, but the cost would also be high, with so many SSDs, even if at that number of SSDs each SSD could be the smallest that can be found with a PCIe 5.0 interface.
Astra-Ultra? Even the largest open model to date (Kimi K3) is nowhere close to Astra level, and it will be quite slow even on the highest-spec M5 Ultra, with achievable speeds of about 0.5 tok/s at most due to having to stream weights from SSD (~13 GB/s on the highest storage capacity M5 Max machines so far). This is OK for doing simple Q&A in the background but it's far from a genuine coding experience. You'd have to test batching of multiple thinking streams in order to try and raise overall tok/s via layer-wise reuse of the streamed weights (and this is where the "Ultra" part sort of becomes relevant; Kimi series models have good support for agent swarms) but this would decrease single-session performance even further. It would only be usable for background jobs, though the hardware would then have a chance of paying for itself if it was fully used on a 24/7 basis.
That's all well and good but Qwen 27B is a small, dense model; that's favorable to both batching and MTP. Batching of large, sparse/MoE models like Kimi K3 (requiring slow SSD streaming even on a single maxed out Mac Studio) on local hardware is an entirely different game that's mostly theoretical so far: many people would even call it outright pointless. (MTP clearly fares even worse, though - unlike batching, it ends up wasting scarce weights-fetching throughput on wrongly predicted tokens.)
Parent commenter hinted at that. Yet DeepSeek has released their V4 which was hugely successful, and even their new architecture is marked V4.1. Qwen internals mark their Flash-Next model, also very compelling, as "qwen4exp". So both of them are bucking the negative stereotype.
> possibly done via the Manim library or something like it
It can be done already: the point is that the motivation and explanation parts are terrible, especially for novel topics where the AI can't just rip off existing content. A Lean proof is at least a verifiable task; you end up with an actual proof that you can work through. A Manim slop video doesn't have that.
reply