Basically the title.
For me, a dense model around 20b, with an additional engram table of 10-20b, or a sparse MoE around 80b would be amazing.
You must log in or register to comment.
gemma 26b-a4b was fantastic, anything in that area is ideal
running my tiny rx6600 8gb vram
Oh hell yeah
I’m on my 1070 8gb copium so we’re twinsies
Maybe they release a huge moe model with non-uniform quants and then we can get even more quality on those cute little cards
How’s your experience with the 1070?
Mixed?
It’s the lower bound of where LLMs are helpful. Anything less and the models would be either too slow to work or too dumb. With the 1070 they are already painfully slow but at least they’re still good enough to do some menial tasks.
A good harness for web search etc does a lot though.
I have an M4 Pro 24GB and I still use Gemma 26B-A4B as my primary model. Beats every other model I’ve ever tried locally on most stuff except coding and tool use, so a Qwen 4 at this size would be super interesting
Another 35B-A3B MoE or similar would be nice…
FYI, Yandex released an 80B MoE base model (i.e. not instruction tuned!) recently: https://huggingface.co/yandex/AliceAI-Foundation-80B-A3B-Base – haven’t experimented with it personally though.
I hope an instruction tuned will be released as well.
Flash next 125b is genuinely amazing, I hope to see 4.0 of the same size.
Agree, and upvoted.
Another good size would be 160 gb for 96 gb vram systems. After that I would like to see 225 to 250 gb.
32GB RAM and 8GB VRAM here, so 35b a3b is about my maximum, maybe with added n-grams
27B dense was a good size for 3.8, but I’m guessing they’re going to add a few B parameters in order to make it noticeably better.
Not necessarly, I’d say. We’ve seen many releases with multiple sizes and the 27B class has been around since at least the Gemma3 times.
Well, it’s just a guess. But also Gemma4 increased to 31B, so I think there’s a possibility that Qwen does the same.
If they can make the KV cache size scale more efficiently with many tokens, like DeepSeek 4.1 Flash fitting 1 million tokens in 890MB, it might give them enough room to add a couple billion parameters of weights and still fit well in 32GB GPUs
They’ve changed the license with Qwen 4, starting with 3.8-next. So I probably won’t be using any further releases
What are the significant changes? Requiring attribution or something more sinister?
Something that would make good use of a 128GB spark. GPT-OSS-120b is good but shows its age
30-35B or just under seems perfect to me.
