Introducing Decision 3.0: towards open multimodal foundation decision models ⚡ Open and state-of-the-art from 0.6B to 27B: #1 on both Jev Decision Index boards in our evaluation, on a new Pareto frontier. HF collection 🤗: #OpenSource #Jev #LLM

Congratulations! Would love to get Decision 3.0 models on EdgeEvals clinical system one benchmark 🔥

Thanks Maziyar!

🤗

All sizes of d3 models published at official jev decision index, and 5 of 6 released models (except the smallest 0.6B , also top2 at the same size), reached at the pareto frontier. This means, no matter which model you wanto to pick, it is always the best at the same size. 🪄

You guys are way too fast, what's your secret? 😏

Just don’t sleep 🤗🤗

🐮

Officially d3 is the new top now! This is the celebration to the open source ecosystem, power of the community and great minds meet together to push the new SOTA 🤗

A single recipe holding the top spot from 0.6B to 27B is the real result — most releases only show that at one scale. We have seen decision models land close to purpose-built rerankers on quality once the request shape is right; good to see it hold at the small end too.

Yes we aim to provide more choices to the community

Open weights across that whole range make the scale-quality curve checkable against the closed Jev board, not just claimed.

Congratulations on the release! Looks super strong!

Open weights for decision models is the right move, but one caveat from production experience: topping a decision index doesn't transfer cleanly. Decision quality is brutally task-specific — the mix of calls, retries and escalations your harness actually makes rarely matches a benchmark's distribution. The teams that win with this will be the ones who build their own eval over their own decision traces, not the ones who pick the index leader. Open weights mean everyone can run that eval locally. That's the real unlock.

0.6B to 27B in one release, my weekly quota still dies at the same spot

Haha 🤗 more to come! 🙇‍♂️

Small open decision models are the quiet half of this story. A 0.6B gate that knows when to wake the big model is how intelligence moves onto hardware people own. A smarter model alone is not necessarily a better intelligence system. The routing is the system.

I like that there are different model sizes to try.

Yes ! 🙌

Impressive scaling from 0.6B to 27B while staying open

Thanks 😊 More to come!

Pushing open multimodal decision models onto a new Pareto frontier like this is a huge win for the community.

More to come ! 🔥

number one on a leaderboard you graded yourself

congrats sir, amazing work!

Thanks : )

Incredible scale and performance love seeing this released

Exciting progress for open multimodal AI. Strong performance across model sizes makes this especially interesting for real-world applications.

Finally seeing open decision models that cover the full size range instead of just one flagship. Beating closed 9B models is nice on paper, but self-reported #1 rankings always need outside verification.

Ahhh？不是才刚刚发布2.0吗？？这今天就3.0了？wtf？🫪

Haha yes 🙌

Open multimodal decision models at this scale is huge news

We need quantizations and a way to serve in prod want to use it for video CCTV alert system for that need batching efficient serving

Huge collection

Impressive open source milestone Scaling from 0.6B to 27B is exciting.

#1 on both boards "in our evaluation" is doing a lot of work in that sentence. got a held-out set, or is this the eval the index was built on?

80% private dataset

@ClementDelangue congrats on the launch. the blender demo shows where this gets real: once a model decides refunds, the next question is which decisions it can make alone and which ones need a human yes. curious if you're looking at confidence thresholds for that

The smaller models might end up being the bigger story here.

Indeed!

A range from 0.6B to 27B is useful, since the right size depends on how fast the router has to be. Which size would you start with for text routing, and how well does the multimodal side hold up on screenshots?

Open multimodal decision models hitting number one is impressive

multimodal decision models at 0.6B is wild, testing this today

Have fun!!

This is awesome. Open, multimodal, and topping both Jev Decision Index boards from 0.6B to 27B is huge. Pushing a new Pareto frontier and keeping it open source too really exciting to see.

outstanding work ! amazing model!

Thanks 🚀

Open multimodal and topping the boards from 0.6B to 27B huge launch

Wow, Vision support is what was lacking from others. Great work, the opens new frontiers!

a family from 0.6b to 27b is the interesting part to me, does a fine tune or adapter made on one size carry to the others or does every size need its own, that seems like the missing piece for model families

Hitting #1 across that range open source is impressive

Decision 3.0 sounds groundbreaking, open multimodal models from 0.6B to 27B topping Jev boards is hugely impressive work