3 ms·

Laguna S 2.1 is really great too, in the "preview" release they've done so far at least. Still pending some reasoning-looping, but besides that, it's a really s

Laguna S 2.1 is really great too, in the "preview" release they've done so far at least. Still pending some reasoning-looping, but besides that, it's a really strong model to run within 96GB VRAM with the NVFP4 variants, and it's really good at coding (specifically).

kadoban 2mo ago
Yeah I think it got bad press because the chat templates (or something?) were messed up on first release, but I've been using a quant of it and it's a powerhouse, better than qwen 3.6 27b for local on a 3090, which is saying a lot.
- embedding-shape 2mo ago
  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.
behnamoh 2mo ago
No it doesn't follow instructions and is substantially slower than ds4.
- jauntywundrkind 2mo ago
  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://earendil.com/posts/session-portability/ https://news.ycombinator.com/item?id=49118781 https://news.ycombinator.com/item?id=49118781
  - behnamoh 2mo ago
    I like the transparency of its reasoning, and I agree with you, OpenAI/Anthropic/Google should show the reasoning traces as well.
- kadoban 2mo ago
  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.
  - Tepix 2mo ago
    Are you talking about S or XS? S is too large for a 3090 at 118b parameters.
    - kadoban 2mo ago
      S, I'm running one of the unsloth quants. Q4 something K_XL I think? I don't think it's all resident in memory on the 3090 but it runs fast enough for use, I'd have to check but it's like 30 t/s or so, fast enough that it doesn't bother me.
  - ericd 2mo ago
    I found Laguna S to be pretty good at coding, pretty fast, but pretty bad as an agent - not proactive, would frequently stubbornly argue things that weren't true, and pretty bad general knowledge. But as a pure coding model, pretty good. Deepseek v4 Flash 0731 is so much better if you can run it, though. Grain of salt, I think I grabbed Laguna after they fixed the initial looping issues, didn't notice those, but there might've been other fixes since.
    - embedding-shape 2mo ago
      > But as a pure coding model, pretty good. Yeah, this is my perspective too on Laguna S 2.1. Works amazingly for coding, pretty bad for pretty much anything else. I don't do a lot of advanced math, supposedly it's good for that too.
- SwellJoe 2mo ago
  It is not slower than DS4. That's crazy talk. It's much smaller, with smaller active parameters, and thus runs faster.
walrus01 2mo ago
There was an obvious problem in the original release, they re issued it after like a week with the reasoning looping supposedly fixed.
- embedding-shape 2mo ago
  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.

kadoban 2mo ago

Yeah I think it got bad press because the chat templates (or something?) were messed up on first release, but I've been using a quant of it and it's a powerhouse, better than qwen 3.6 27b for local on a 3090, which is saying a lot.

embedding-shape 2mo ago
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.

behnamoh 2mo ago

No it doesn't follow instructions and is substantially slower than ds4.

jauntywundrkind 2mo ago
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://earendil.com/posts/session-portability/ https://news.ycombinator.com/item?id=49118781 https://news.ycombinator.com/item?id=49118781
- behnamoh 2mo ago
  I like the transparency of its reasoning, and I agree with you, OpenAI/Anthropic/Google should show the reasoning traces as well.
kadoban 2mo ago
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.
- Tepix 2mo ago
  Are you talking about S or XS? S is too large for a 3090 at 118b parameters.
  - kadoban 2mo ago
    S, I'm running one of the unsloth quants. Q4 something K_XL I think? I don't think it's all resident in memory on the 3090 but it runs fast enough for use, I'd have to check but it's like 30 t/s or so, fast enough that it doesn't bother me.
- ericd 2mo ago
  I found Laguna S to be pretty good at coding, pretty fast, but pretty bad as an agent - not proactive, would frequently stubbornly argue things that weren't true, and pretty bad general knowledge. But as a pure coding model, pretty good. Deepseek v4 Flash 0731 is so much better if you can run it, though. Grain of salt, I think I grabbed Laguna after they fixed the initial looping issues, didn't notice those, but there might've been other fixes since.
  - embedding-shape 2mo ago
    > But as a pure coding model, pretty good. Yeah, this is my perspective too on Laguna S 2.1. Works amazingly for coding, pretty bad for pretty much anything else. I don't do a lot of advanced math, supposedly it's good for that too.
SwellJoe 2mo ago
It is not slower than DS4. That's crazy talk. It's much smaller, with smaller active parameters, and thus runs faster.

walrus01 2mo ago

There was an obvious problem in the original release, they re issued it after like a week with the reasoning looping supposedly fixed.

embedding-shape 2mo ago
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.