Lumina

AI News | Benchmarks

[Lumina @LuminaBench]

🚨Muse Spark 1.2 Leaderboard Rankings

I’ve added it across the main LuminaBench categories here are the main ones:

• Overall: 78.5
• Coding: 77.0
• Reasoning: 77.0
• Agents: 79.9

Breakdowns for each below👇

https://t.co/Ibhjm4bGvp

[Lumina @LuminaBench]

@Lucas09aoht don't think it makes a difference tbf but i normally default extra high on web

[Lumina @LuminaBench]

🚨OpenAI is A/B testing image generations still

I just got this directly inside ChatGPT, two different generations from the same image prompt.

This comes right after (mona-lisa-1) appeared on Arena, so a new image release could easily come this month.

I’d already caught OpenAI A/B testing image upgrades before this too, there is a massive upgrade in text generation from what I can tell.

https://t.co/qinfHgwo8p

[Lumina @LuminaBench]

🚨A new OpenAI image model just surfaced on Arena

It’s showing up under the codename “mona-lisa-1”, and looks to be a new unreleased OpenAI image model.

This could be what OpenAI has been A/B testing over the past few weeks, and it looks like more than just a small upgrade.

https://t.co/sdFhUOrQ4R

[Lumina @LuminaBench]

@ajs6888 yup

[Lumina @LuminaBench]

GPT 5.6 Sol Pro is incredible, but the limits are ass

The model is one of the best I’ve used for turning my ideas into perfect prompts for codex, but you burn your usage in an hour or two with it.

The limits desperately need upping.

[Lumina @LuminaBench]

@polash_ai Yeah, I want to see how qwen 3.8 27B competes on Wednesday now

[Lumina @LuminaBench]

🚨 Meta just open sourced Muse Glimmer

The weights are now available for their 30B dense model, small enough that a lot of people can actually run it locally which is nice.

And more importantly, Muse Spark 1.2 weights are coming soon too.

Spark 1.2 is already a ridiculously strong model for its price, so getting the full weights for it will be very interesting.

https://t.co/zSxu8JccS9

[Mark Zuckerberg @finkd]

Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.

[Lumina @LuminaBench]

@ajs6888 It’s an open model that is meant to be run locally

[Lumina @LuminaBench]

@gurbakshchahal @thsottiaux Yup

[Lumina @LuminaBench]

what's the plan if @thsottiaux doesn't reset limits today?

https://t.co/yfBlDXwUYP

[Tibo @thsottiaux]

@rxmphai I'll do another performative reset on Monday

[Lumina @LuminaBench]

@PJeevanG1 @thsottiaux I ran max/ultra all sunday, my usage is like 10% left on my pro plan 😭

[Lumina @LuminaBench]

@TabDuoBao most AI message ever

[Lumina @LuminaBench]

@Sidlik900 yeah they are getting much better too

[Lumina @LuminaBench]

🚨Qwen3.8 27B drops next Week

This might actually be a more interesting Qwen release than the 2.4T Max model:

• 27B parameters
• Open weights
• Drops August 12th
• This should also run locally on around 17GB RAM/VRAM when quantised, so a very low bar to entry

The 20 to 30B Parameter model range is getting very good, what you can actually run locally now compared to even 6 months ago is a huge jump.

Really excited for this one.

[Lumina @LuminaBench]

@rob_x_ai 👀

[Lumina @LuminaBench]

@minkrov they are getting closer, text was notoriously bad and is improving very quickly now, check out the qwen release specifically the text image on it, massive improvement
https://t.co/r9VikmwnGK

[Lumina @LuminaBench]

🚨 Alibaba just launched Qwen-Image-3.0 and it looks ridiculously capable

The official improvements include:

• Up to 4.5K-token prompts
• Legible text as small as 10px
• Native rendering across 12 languages
• More than 100 supported visual styles
• Complex newspapers, exam papers and interfaces generated in one pass
• Entire 3×3 grids containing nine detailed infographics

There are no independent benchmark results yet, but the examples Qwen have shown are incredible, I can't tell if some of the images are AI or real, they may have just built one of the most useful image models available.

The Chinese labs are moving unbelievably fast.

Do you think this beats the competition in image gen?

[Alexandr Wang @alexandr_wang]

personal superintelligence should be available to everyone, and opening access to our models is abig part of that. read more from mark: https://t.co/ij03Aubt1D

[Lumina @LuminaBench]

@TJeparskis59091 @SemiAnalysis_ it won't get released in august, more like october onwards most likely google are really slow

[Lumina @LuminaBench]

🚨Gemini 3.5 Pro might have been cancelled

SemiAnalysis is now claiming Google silently cancelled Gemini 3.5 Pro and shifted its attention towards Gemini 4 instead.

I swear to god if after all those months of waiting they just cancel it im done with google, never again😂

[Lumina @LuminaBench]

@FayLani0 💀

[Lumina @LuminaBench]

@iruletheworldmo dougtrio 😂

[Lumina @LuminaBench]

@tembakhumalo @SemiAnalysis_ yeah my student one ends in september i have used it like 5 times in the year i have had it

[Lumina @LuminaBench]

@minkrov parameter count doesn't always make it costly, OpenAI are making a big push for cost efficiency too, i wouldn't be surprised if it keeps the same price as Sol roughly, or if they go hard with post training they can make it much cheaper

[Lumina @LuminaBench]

🚨Astra leaks are getting ridiculous

Quick correction first: Doug is apparently the codename for Astra’s pre train not a new one.

Latest claims:

• Astra will be around 10T parameters
• Built from the new Doug pre train
• Supposedly it crushes Fable across benchmarks👀
• Apparently a huge jump in writing quality from it too
• Still expected this month even with the setback
• And there’s an even bigger pre train leaked for the end of the year

I’m expecting this to be the first model that genuinely moves past Fable/Mythos and is a definitive leader across most benchmarks.

https://t.co/S84mEjjWJ5

[Lumina @LuminaBench]

🚨GPT 6 might not even be OpenAI’s biggest model this year

A much larger pre train is leaked to be coming around November time:

• It is Codenamed “Doug ”
• I think this is the model OpenAI alluded to in June
• This is also going to be OpenAI’s biggest pre train yet
• It is also supposedly strong enough to make Fable seem “primitive”😭

I genuinely can't fathom what this model will be capable of if Fable is primitive to it, this thing is going to be ridiculous.

https://t.co/Qzc4vKxWNa

[Lumina @LuminaBench]

@shadowaguy you could be waiting a while

[Lumina @LuminaBench]

@mPanda3D 😂 not far off