No, the quants they released were also messed up. RC2 also ballooned the size so the ones who were excited about RC1 (like me) can no longer fit it in our hardware. They haven't promised anything, but said they'll try to restore the RC1 size for the next update of the weights.

Like glm-5.x I think it has enormous self introspection that it often trips up on, but that this self reflection is actually a superpower, that enables incredibly good output. And from (in some cases) very small models.

If you watch it think, which you can, unlike American closed models, you can steer it. You can provide a a massive rocket ship stratospheric boost to help it orient itself. You have no self correction, there is no multiplayer in American proprietary models.

Sure it's great having super powerful mystic oracles that have the "right" answers. But I love respect & revere the open thinking. No it's not automous. But it is brilliant. And it considers. A lot. Deeply. It chases. That to me is the most human of models, even as it falls far astray.

You should help it. You can. Unlike these vicious dark surfaces which yield and tell you nothing. I think this is the actual meta-core-super-point of "The session you cannot take with you" (link below). It's the session that does not care about you, will not interact with you, will not peer with you, that is a dead remote far off oracle to you. Fuck these "oracles". They are a plague against the human spirit. We should alloy humanity and AI to Augment Intellect (Engelbart). (To do less is species treason.) https://earendil.com/posts/session-portability/ https://news.ycombinator.com/item?id=49118781

It's a lot smaller, and runs (quantized) on a 3090 quite well. Ds4 flash 0731 you're talking about? It's great but it's much harder to run locally.

It is not slower than DS4. That's crazy talk. It's much smaller, with smaller active parameters, and thus runs faster.

Well, bit more complicated than that, I've been eagerly helping in testing and keeping track of what they've done. Initially there were serious bugs, also about the templates, eventually they released RC1 which had some fixes towards the looping. Then a couple of days later, they released RC2 which supposedly fixed the issue, but ballooned the size so all of us who were running Laguna S 2.1 on a single Pro 6000, suddenly could no longer. So, unsure if RC2 actually fixes the issue, as we're a bunch who can no longer run it :)

Besides that, it was also discovered that their suggested inference parameters were wrong and led to worse behavior. Eventually someone discovered these works best (so if you have the issue with looping right now, try these, helps a lot for me but not 100% still) and was also what the evals used apparently: temperature: 1.0, top_p: 1.0, top_k:20

Now we're waiting for RC3 which Poolside said will come at one point, and hopefully also brings down the size again NVFP4 weights + full context can load properly again even on "smaller" hardware.

Nemotron 3 also introduced LatentMoE, which was adopted by Kimi K3 :)