VARIEN

composer, art director, digital forensics engineer
spreading the joy of open source technologies
God’s in his heaven—All’s right with the world

[VARIEN @varien]

We've open-sourced Qwen Concept 27B, our first model.

It applies direct weight-space edits (inspired by the recent MIT CSAIL "Sparks of Weight-Space Programming" paper) to cut thinking tokens by 54% while improving GPQA Diamond performance by 3% versus the base Qwen model.

This yields 30% more useful tokens per joule with no additional post-training, all developed on local NVIDIA DGX Spark hardware.

BF16, FP8, and NVFP4 versions will be available free at https://t.co/ZdZGxIEIbg.

[Concept @conceptcountry]

Concept Country has developed its first open-source model, Qwen Concept 27B.

With 54% less thinking tokens and 3% better performance on GPQA Diamond compared to its base Qwen 3.8 model, we have achieved 30% more useful tokens per joule without additional post-training.

[soli @solisolsoli]

Malenia, Blade of Miquella

[VARIEN @varien]

@somewheresy Much was learned. But I do give it 2 years or less until we can see something compelling longer than 10-20min. My guess is SD 3.0 gets 60s generation time which, prompt coherence gets stronger, etc. 2027

[VARIEN @varien]

@somewheresy its incredibly clear from his background and The Technological Republic and his time at Frankfurt School and European uni that he's grey tribe (also he can say It)

[VARIEN @varien]

@Ag0at Bread Prison sounds like a Cruelty Squad section

[VARIEN @varien]

"I'm never gonna be my parents" people when it's time to be worse

[Bones @FrailSkeleton]

Business idea: child jail. You drop them off for like a half day when they’ve been really bad. They have to wear the striped clothes& sit in cells, maybe break some rocks. Bread& water only for snacks

[VARIEN @varien]

Writing about this

[ned! @nedsore]

The most mentally ill person you know is currently training to be some sort of therapist

[VARIEN @varien]

@ArtificialAnlys I'll give it a go!

[VARIEN @varien]

@IceSolst That's just CVE-1

in the beginning
Horny

[VARIEN @varien]

@kojinglick 1000 Year Dontonburi Flourishment

(Time to open a cabaret club or get into real estate)

[VARIEN @varien]

@_ontologic Have not heard of this band

[VARIEN @varien]

@jtevesobs So very deserved!

[VARIEN @varien]

@nahid_pro09 💋

[VARIEN @varien]

@somewheresy im in Shenzhen gluing parts together and huffing fumes, truly american

[ComfyDevs @ComfyDevs]

Two weeks ago we asked you to open source Higgsfield.

Then they tried to shut us down. 300+ of you shipped anyway.

Submissions from 45+ countries and 3 build nights later…

The winners of #OpenSourceHiggsfield

🥇 Best Overall: ComfyUI Panorama Stickers by @noma_door
💡 Most Innovative: Blocksfield by @SquaredCubeRBX
🔥 Most Viral: Imagination Studio by @SamJWasserman & @u1tra_instinct
🎨 Open Weights Only: Paint a Plate by @AInVFX

Huge thanks to @runpod_io, our partner and cash prize sponsor.

Plus 10 honorable mentions, with bonus prizes from @ltx_io, @bfl_ai, @runpod, and @MiniMax_AI:

Reshoot 4D by @itsmetamike, OneShot MV by @VisualFrisson, Freefield by @neuregex, H3 Motion by PxTicks (YT), Splat Edit Engine by @TheBurgstall, Sonder App by @SonderSaid, Max Editor by jonathan_7gramirez (IG), Genobi by Radek Rudnicki (LinkedIn), Henge by @Alissonerdx, and H3-AudioRetake-Studio by @nebsh83.

To everyone who built: thank you. You didn’t just enter a challenge. You showed that the best creative tools don’t have to sit behind a closed platform. Every project here was built in the open and shared so others can build on it.

Built by the community, for the community.

Creativity stays open source.

A special thank you to our partners as well.

Stay Comfy. 🤝

[VARIEN @varien]

@somewheresy oh good to know right before i come

[VARIEN @varien]

@robinjhuang @ComfyDevs King

[VARIEN @varien]

@heroictripod297 Yes

[VARIEN @varien]

Qwen Concept 27B: we rewrote 9 weight matrices in Qwen3.8-27B. No gradient steps, no training data.

It thinks less and scores higher: GPQA Diamond +4.3 pts // LiveCodeBench v6 +13.0 // 27–54% fewer thinking tokens.

Open weights and methodology paper below↓

[Concept @conceptcountry]

Weights and working paper are released. As we run more robustness checks, we also discover that this model edit unexpectedly adds +13% performance on coding tasks.

https://t.co/9TkSUEDTtE

[VARIEN @varien]

all of this sucks just use llama.cpp or @actualinc e

[Uzair Afridi @uzair__afridi]

Can I ask a dumb question?

If vLLM is so much faster than Ollama, why does everyone still use Ollama?

[VARIEN @varien]

@ilblackdragon @somewheresy

[VARIEN @varien]

I hope digital artists, cinematographers, concept artists, set dressers, vfx crews, game devs, etc. are prepared for the next few years.

Listen, you have the ability to construct a plan for AI assistance in the arts, expand your domain of possibilities, or you can dig yourself into a corner, resentful and irrelevant.

I'm saying this to be mean. I am saying this because it has been happening, and is only accelerating.

Good luck and have fun! 🙏