Introducing Decision 2.0: our newest state-of-the-art decision models, open in every size from 0.6B to 27B ⚡ HF collection 🤗: #OpenSource #Decision #Jev #LLM

Wow, covering 0.6B to 27B sizes, all open sourced. This is really solid work.

Thank you!

@airesearch12 to add in the jev bench please

@airesearch12 Let’s do it!

0.6B to 27B. the dropdown is where model choice happens.

Exactly!

I like that you included 0.6B. That's the size I'd try first for a local setup.

Exactly! 💯

Is there a Vision model planned?

Yes for sure : )

I can't wait to try it!

The per-option probabilities are a useful touch. Have you measured calibration after rephrasing or reordering the choices? I'd use that to set a fallback threshold; a harmless label edit shouldn't quietly change which cases reach the larger model.

Good point. Cascade for decision model will be added into to have realtime upgrades

Have you benchmarked it on predictive benchmarks like BeyondArena?

Kai is particularly interesting as it can be used on an edge device

Yes, it is very cheap and capable for most of works, and we can cascade to larger one when uncertain

@cristian_le0 more interesting decisions will come not just vary from the model size 🤗

We benchmarked all six vLLM Decision 2.0 models on StreamDecisionBench, across presenter control, live debugging, assembly coaching, and call assist. Lux-9B came out strongest on the in-force score, while Vega-27B showed strong untimed accuracy but dropped much more once latency was counted. Code and full results are here

Curious what the licence looks like on the big one.

This is a great will for sure will be intergerating this into our own software decision making is the biggest problem these frontier models still lack in great work!!

I see this is based on Qwen. Is it possible to have the model loaded in memory serve decision jev and generation requests?

hey @UnslothAI @ollama can we add decision 2.0 into the stack?

Fully open-weight from 0.6B to 27B is what makes this category actually buildable — teams can run decisions locally, fine-tune on their own judgment patterns, and keep data in-house. The spread across sizes means you can pay for scale only where decisions are hard. What does Decision 2.0 do differently on calibration vs Kev and the other open decision models?

Hey @grok bu nədi belə hər kəs model çıxardır bu neçənci dəfədir görürəm bu gün axı çətin idi hesablama gücü və S. Staratplar şirkətlər ən böyüklərə dirənən bilmirdi

I can't wait to try these!

In this world we course need more decision models

great job 😍

Cool, I like open source

0.6B to 27B, every size. the small end matters more than the headline 27B for on-chain stuff where you just need a fast flag, low latency beats raw accuracy when you're scanning thousands of contracts.

Thanks 🍻

0.6B to 27B open. running the 0.6B locally to grade orderbook snapshots - 3.46B active on Kolibri is still my classifier, but a decision model that fits on one card is worth the swap test. what's p99 on the small one?

Is this vision ? By the way is there any decision model that supports vision right now?

The 0.6B end is the one I'd grab first - big enough range to test on a laptop before committing.

decision-specific training beats general purpose models every time for this use case

0.6B to 27B all open is the useful part. small one fits a router hook, no endpoint to babysit. did they ship the router config too or just weights?

64 decisions in one 63ms forward pass is the bit i'd benchmark against “just ask the main LLM”. this feels less like a tiny chatbot and more like a vectorized policy primitive for routers/agents

love this

Thanks!

@XunzhuoLiu is that better than laya 421M? Can you compare, im curios about that 0.6 , 0.8 reasoining which one equal to vanilla laya 421M?

Does it has some arch improvements over other decision models to run faster over vllm, or does vllm has also improved the infra to run decision models faster?

We have some runtime specific optimizations

Open decision models across that size range is a useful drop here

Thank you !

I ll try to make one play world warcraft by itself .

Cheers 🍻