35 comments

[ 2.7 ms ] story [ 42.6 ms ] thread
If it's ~30% bigger and not as good as GLM 5.2, why would I tinker with this model?
If they have a really seamless fine-tuning experience and maybe can help you extract the data you need to FT (which is one of the big challenges in actually getting fine-tuning democratized), maybe you would use it because "Tinker" defaults to it.

The model could also be more flexible for non-coding use-cases (they show the results for reasoning being strong) so maybe the argument is to use it for non-coding use-cases to drive relatively deterministic conclusions for non-coding agents (they have also done some determinism work on kernels, which could be useful in pulling on that thread of deterministic models that are fine-tuned for everything that is not writing code.)

That said, I'm not sure how much all the work they have done actually synergizes or if the market size (at least in the short to medium term) is big enough for a huge outcome from the company's current valuation with those bets as the enterprise agent estate is taking a while to evolve. Hence companies like Anthropic and OpenAI are throwing tons of consulting money at the problem.

> If it's ~30% bigger and not as good as GLM 5.2, why would I tinker with this model?

The benchmarks never tell the full story. Some of the open weights models have been benchmaxxed for a while. Their utility on real work can be different than the benchmark number.

The multimodal input is also a big deal. Having vision input is really helpful for a lot of tasks.

Then why are they publishing the benchmarks which makes them look worse than GLM 5.2?
Because it's still informative
I'm not sure why I'm being downvoted but I didn't mean it in a negative way.

For such announcement, I would expect them to give me clues on when I should use this model and in which cases it's the best one.

The benchmarks that they share doesn't indicate that it's cheaper to run than other models, or can fit in my local machine, or excels in a specific vertical.

After reading the comments here and X, I can see it being the top-3 multi-modal open-source model though.

I second that. Gemini 3.5 Flash rocks the benchmark charts but is terrible as an agent. Horrible instruction adherence and makes WAY too many tool calls
which cheap models have you found work best as agents?
Most of the bigger open weight models are pretty good. You can get them per-token from companies like Fireworks or OpenCode
Gemini 3.1 pro is still better at creative writing than any other model.
There's also an Inkling-Small that is 276B, 12B active that is much smaller than GLM 5.2 and still multimodal. Not released yet, but in the announcement link they mention that they're testing Inkling-Small & will release as open weight after testing. That one may be interesting as a Deepseek V4 Flash replacement.
It's nice to see a strong long context open weights model that is multi-modal.

There are many applications that will benefit from the strength in audio here and until z.ai and co work in visual this could be very strong for general agentic applications, though I see there's a bit of weakness in the benches for areas that might make that less true.

Like all models need to slap it in your harness and do proper evals on the tasks you care about.

MiniMax M3 and DeepSeek v4-Pro are highly capable long context open weight multi-modal models. But long-context is a trap, because performance still falls dramatically after 150k-200k context.
> But long-context is a trap, because performance still falls dramatically after 150k-200k context.

I'm not sure exactly what causes the difference, but this heavily depends on the model. In my experience with Opus 4.8, I can go well over 500k and still get extremely good results. A drastically different example was GLM-5.1, which worked great until about 100k and then turned insane almost immediately. They did fix that with 5.2, though.

5.1 going insane was probably also a inference quirk. Because it sometimes remained coherent the entire 200k context length.
> But long-context is a trap, because performance still falls dramatically after 150k-200k context.

I often see this repeated, and it is not true task to task. I work on this daily and we have several tasks where long context is advantageous and our evals against a whole battery of models with different windows show it as being so.

This is why having good evals for the tasks you're working on is so important.

I do grant it's a good rule of thumb.

Raised 2 billion dollars at a $12 billion valuation and debuts at 41 on the Artificial Analysis Intelligence Index. What a joke.
They also indicate they have a 276B A12B version, but it doesn't seem the weights are available. This might actually be able to fit in 128GB when quantized to 2 bits or so which makes it interesting.
Cool, now we just need the GPU that supports it
For the most part it’s better than Nemotron, worse than GLM. This makes it the best American open weights model from what I can tell?
It's nearly double the size of Nemotron 3 Ultra, so I'd expect it to be considerably better, although the active parameter count seems to be a touch lower at 41B vs 55B
I'm surprised that Nemotron gets mentioned at all. In my experiments with it for coding tasks it performed extremely poorly, essentially unusable.
it is pretty good at instruction following and has extremely fast decode.
I focus on realtime voice AI uses cases and nemotron's time to first token is INSANELY fast. It's become a legit option for voice use cases
It looks like HuggingFace shows Apache-2.0 but they have AUP. How does it work together?
(comment deleted)
too bad we'll never know how good it is, since they used a radar plot to show its benchmark scores!
How does the radar plot prevent you from looking at just one of its axes?