Sink

hello world

[Sink @Sink1737182]

underdog invite, anyone? @underdogai

[Sink @Sink1737182]

@underdogai Haven't got my invite yet@underdogdotai

[Sink @Sink1737182]

@underdogai pls invite me

[Sink @Sink1737182]

@pswai just use train data

[Sink @Sink1737182]

@pswai the test data was never trained on .

[Sink @Sink1737182]

@pswai 27B is better！

[Sink @Sink1737182]

@Asteri_eth The three decisions per turn are a good fit for a shared forward pass. We released StartLux-Decision with a Jev-compatible API and local GGUF/MLX backends (text-only). Weights are CC BY-NC 4.0: https://t.co/ILzRtqsKCp

[Sink @Sink1737182]

@rayanabdulcader We released StartLux-Decision for these loops too: per-option probabilities, image input, and weights you can run yourself. GGUF is text-only for now; weights are CC BY-NC. Game demos + code: https://t.co/ILzRtqsKCp

[Sink @Sink1737182]

@0xSigil @im_roy_lee @underdogai Invite plz sigil

[Sink @Sink1737182]

@shiftb @underdogai Pls send me an invite

[Sink @Sink1737182]

@shiftb @underdogai Can I get an invite?

[Sink @Sink1737182]

@NaceAI https://t.co/ILzRtqsKCp

pls look this repo：Decision Index 0.2.1 and jevbench

[Sink @Sink1737182]

Code, eval scripts and raw results: https://t.co/ILzRtqsKCp
Weights: https://t.co/zWr12UsGXO

If it's useful, a ⭐ on GitHub or an upvote on Hugging Face would mean a lot. Thanks!

[Sink @Sink1737182]

We just open-sourced Startlux-Decision, decision models from 0.8B to 27B.

On Decision Index 0.2.1 we measured 63.88 for the 27B and 58.63 for the 9B . The top public entry is Jev 1.13 at 57.91.

On Intern-Decision's seven suites, the 27B averages 91.82% (Jev: 88.74%).

[Sink @Sink1737182]

Demos: it orders from a web store on its own, clears Super Mario Bros. 1-1, and makes the strategy calls for a StarCraft II bot that beats the built-in Hard AI.
GGUF for llama.cpp too: the 4B Q8_0 (4.48 GB) matches the original on all 231 public JevBench items.

[Sink @Sink1737182]

@CoderDaMing @grok 把上面整理为英语发给我

[Sink @Sink1737182]

@TakeProfitLLC @goated_fx
@TheEntryExit
@Tejsviraghav

[Sink @Sink1737182]

@elonmusk cool

[Gradient @Gradient_HQ]

Intelligence has been locked in walled gardens. Today, we’re opening the gates.

Parallax now runs in Hybrid mode, with Macs and GPUs serving large models together in a truly distributed framework.