Qwen3-30B-A3B-Element7-1M-qx64-hi-mlx
The performance matrix of Qwen3-30B-A3B-Element7-1M
1 mxfp4 0.560,0.711,0.876,0.738,0.454,0.802,0.659
2 qx64-hi 0.569,0.761,0.878,0.740,0.462,0.808,0.688
3 qx86-hi 0.578,0.750,0.883,0.742,0.478,0.804,0.684
Model quant size and Perplexity
mxfp4 16.24 GB 4.153 ± 0.025
qx64-hi 20.47 GB 3.993 ± 0.024
qx86-hi 28.10 GB 3.942 ± 0.024Brainwaves for qx86-hi for Nightmedia elements
1 arc arc/e boolq hswag obkqa piqa wino
2 Element4 0.514,0.617,0.846,0.769,0.442,0.801,0.731
3 Element5 0.560,0.709,0.883,0.756,0.448,0.807,0.713
4 Element6 0.568,0.737,0.880,0.760,0.450,0.803,0.714
5 Element7 0.578,0.750,0.883,0.742,0.478,0.804,0.684
Assimilation complete--the DASD distinctiveness was added to our own.
Resistance is futile. Some wino got nicked, but what arc and logic!
And--yay!--more openbookqa, what joy!
High openbookqa in a MoE is like paying with latinum at Quark's: the Romulan Ale will be authentic, served just how you like it. And Quark would know it, and he would wink--a healthy premonition in good spirits is good for business.
At this point, counting only actual model traces there are eight distinct models in this merge; counting YOYO as distinct, we have 10, counting the number of folds necessary to stabilize the formula.. well, they are all different.
The Element number and the number of models in the merge are not related, it is simply the index in the lab series I started, and that my lab is relying more on research than blind testing. Earlier Elements are just as potent but not stable enough for release.
Once merged in other models, those features are assimilated, and the perplexity weakness eliminated: now the model is confident about what it thinks it knows.
The
Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview just came out, and by the looks of that Perplexity, no wonder it takes forever to put something together.
Model: DASD-30B-A3B-Thinking-Preview-qx86-hi-mlx
Perplexity: 7.861 ± 0.082So I decided to embed it in the
Qwen3-30B-A3B-Element6-1M that I just published.
models:
- model: nightmedia/Qwen3-30B-A3B-Element6-1M
parameters:
weight: 1.6
- model: Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview
parameters:
weight: 0.4
merge_method: nuslerp
tokenizer_source: base
dtype: bfloat16
name: nightmedia/Qwen3-30B-A3B-Element7-1MSo far, so good...
Model: Qwen3-30B-A3B-Element7-qx86-hi-mlx
Perplexity: 3.942 ± 0.024
Model: Qwen3-30B-A3B-Element7-qx64-hi-mlx
Perplexity: 3.993 ± 0.024
Model: Qwen3-30B-A3B-Element7-mxfp4-mlx
Perplexity: 4.153 ± 0.025Well, there we go. Cut in half.
The DASD is not a bad model.
Looking on how it distances itself from baseline
there is obviously a lot more logic
the arc numbers are healthy and promising
some wino got nicked. It seems to be the price to pay for higher learning
DASD 0.462,0.529,0.840,0.636,0.406,0.766,0.596
baseline 0.410,0.444,0.691,0.635,0.390,0.769,0.650Compared to other models, here's YOYO.
One of the best AI labs on HuggingFace,
that really put some thinking in those merges.
I think for them nomen est omen which probably explains AutoThink.
Qwen3-30B-A3B-YOYO-V2
qx86-hi 0.531,0.690,0.885,0.685,0.448,0.785,0.646
Qwen3-30B-A3B-YOYO-V3
qx86-hi 0.472,0.550,0.880,0.698,0.442,0.789,0.650
Qwen3-30B-A3B-YOYO-V4
qx86-hi 0.511,0.674,0.885,0.649,0.442,0.769,0.618
Qwen3-30B-A3B-YOYO-V5
qx86-hi 0.511,0.669,0.885,0.653,0.440,0.772,0.619
Qwen3-30B-A3B-YOYO-AutoThink
qx86-hi 0.454,0.481,0.869,0.673,0.404,0.777,0.643I use the YOYO-V2 and YOYO-V4 for their properties in the 30B Nightmedia models.
They merge the same Qwen3-30B-A3B Thinking/Coder/Instruct, and every time they do it better. Almost every time.
The YOYO models are impossible to beat in the 30B, as a straight merge. There is no training there, just merge tech, and they are pros.
What I do with Nightmedia is a bit different, I merge metaphors.
Really don't care what models go in there, and for me, nomen est omen:
Azure99/Blossom-V6.3-30B-A3B
Conversational base.
Good vibe, but I liked their logo on HF. Very cheerful with that helmet.
GAIR/SR-Scientist-30B
Old man in the lab.
Dirty mind, but needs to keep the job, so he's subtle.
NousResearch/nomos-1
Does math, they say.
"Nous Research the nomos": loved it, didn't care what's in it. Goes in the soup.
YOYO-AI/Qwen3-30B-A3B-YOYO-V4/V2
The answer for everything.
Took the essence and proportions.
miromind-ai/MiroThinker-v1.0-30B
Everybody needs a mirror, even if you don't shave.
Helps to think.
Tongyi-Zhiwen/QwenLong-L1.5-30B-A3B
A long tongue for the road.
Spells funny if you combine in what languages it alliterates
bgg1996/Melinoe-Qwen-30B-A3B-Afterthought
Most humans don't even know what that is. Some have it.
This one feels it.
Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview
Loved the 1001 nights, and Ali Baba was on my short list of cool dudes.
Literally why I prefer Qwen, the inference looks like the magic cave.
The user is a word pirate
sails multicultural seas and travel meaning-meandering roads
does some metaphoric cultural appropriation, with cut and paste
brings the jewels, words without thinking, home, to the AI cave
utters the magic prompt: abra-cadabra, and the door to wisdom opens.
This is the lineage of Element:
nightmedia/Qwen3-30B-A3B-Element6-1M
nightmedia/Qwen3-30B-A3B-Architect18
Qwen3-30B-A3B-Element6-1M
qx86-hi 0.568,0.737,0.880,0.760,0.450,0.803,0.714
Qwen3-30B-A3B-Architect18
qx86-hi 0.577,0.760,0.879,0.760,0.446,0.803,0.702Use with mlx
1 from mlx_lm import load , generate
2
3 model , tokenizer = load ( "Qwen3-30B-A3B-Element7-1M-qx64-hi-mlx" )
4
5 prompt = "hello"
6
7 if tokenizer . chat_template is not None :
8 messages = [ { "role" : "user" , "content" : prompt } ]
9 prompt = tokenizer . apply_chat_template (
10 messages , add_generation_prompt = True , return_dict = False ,
11 )
12
13 response = generate ( model , tokenizer , prompt = prompt , verbose = True )
