Qwen3.8 Max is getting open-sourced tomorrow and before it does I want to state my opinion on the model after rigorous use. Although I am a Qwen ambassador this is my full real opinion. The model from my experiences acts more like Opus than it does Sol, it's less programmatic and has more a expressive personality. On lower thinking modes (low, medium) this expressive-ness is very clear, it's more opinionated, moody, and definitely more likely to make irreversible mistakes. Qwen3.8 has been my daily driver for a couple weeks now and I want to say the output in omp is some of the best I've ever got, Qwen3.8 + OMP > GPT-5.6 Sol + Codex, this is a hard truth to straddle as I send hundreds of dollars/month to OpenAI for codex, but it's ability to hold this harness by the neck and work like a dog is truly unbelievable. it orchestrates sub-agents really well, and although a (very large) token burner it never disappoints. However, it still falls far behind in real-world developer experiences and definitely behind Sol/Fable in "out of the box" thinking. This thing is the biggest work horse model I've ever tested and can be set on a task to run for days without any interruptions (they really didn't lie there), to the point where it can spend hours doing extremely deep research on a indirect task. Also, dont use xhigh thinking, waste of tokens.

while testing this I saw there is a new model in codex CLI "Daybreak Blue" It seems to be currently unavailable for the codex app yet I have access through the CLI. Seems like a similar model to Mythos relative to Fable. Fully unrestricted? anyone else have any info?

has anyone extensively used paseo? its a codex-looking UI which just uses codex, opencode, omp, claude.. etc as a back-end. Very similar to the codex app but doesn't drain resources nearly as much. Also has relay + IPhone app which allows connection like codex. The main reason I still pay $200/month for codex is because of the app and integrations, I would switch in a heartbeat to this if it treats me well. along with if anyone has any other recommendations of apps similar to this please lmk

If you want to run local AI for inference, buy one of these. This is an idea I had long ago but never had the balls to bring it to fruition. Really exited with what lucebox has been cooking, really want to get my hands on one of these machines.

benchmarks of a 50% pruned Qwen3.6-35b-a3b and expert-specific quantization technique (made by me) 7.3gb model preforming => 51gb model, exiting to see where I can bring this technique to. I have some more things lined up too. I need a DGX spark😭

Timelapse #14#; 7 hours (9:00am-4:00pm) What I did: - Started experimenting with quantization aware REAP - Studied online post-training techniques (DPO, PPO, GRPO) - Tested qwen3.8 through ambassador program - benched qwen3.8 @neutralityproject - Made 2 synthetic datasets for left/right output of LLM’s - Stretched every hour + posture excises and rice bucket training - Watched rekraps newest video - Watched Martin Shkreli stream - Scrolled X Song: Memory limitations in artificial intelligence - Infinity Frequencies

really exited to see what qwen3.8 has to offer, all the posts saying "its not even close to fable" than showing it merely being bad at 3d design is not an accurate representation of the model. currently waiting on fixes through the dev ambassador program, once fixed I'll benchmark it through @neutralityorg + run it personally and quote with my actual thoughts.

Kimi K3 is bigger than the Deepseek moment. this one release is going to change the world as we know it. At some point models get so good they can do full training runs e2e, 5.6 Sol already post trained Luna, getting Kimi K3 alliterated in BF16 will be a super weapon. full OSINT capabilites, research new drug discovories without being blocked, the list goes on. On top of all of that Kimi K3 will be the perfect distillation model for fine-tunes of smaller models, it will be able to generate amazing synthetic data, all that can be run on rented hardware on 8xB200

We've just merged its first community contribution! 🚀 6 new models benchmarked and submitted via open PR: DeepSeek, Kimi K2, Hunyuan, Nova Pro, Command A, Jamba and two new countries on the map. 24 models total, from 6 countries. And because trust is the whole point: community runs are displayed flagged as "pending re-verification" until we reproduce them ourselves. Open methodology means anyone can contribute, and nobody, including us, gets taken on faith. Thank you bricepirard-spec. Who's next?

When AI embeds itself in decision-making, who oversees its influence? Right now: nobody. We vows to build and open-source the methodologies, datasets and benchmarks that change that. Nothing stays behind closed doors. Helps us fund the research:

after many years on this platform I have finally hit 25k followers. many brainrot posts, ragebait, and genuine public releases have led up to this milestone so far. Thank you everyone, this means a lot.❤️

ok so i am waking up to @elonmusk retweeting our discovery 3 times. thanks! for anyone curious about how we found that Grok is the most neutral model out there, please give a follow to our new initiative @neutralityorg and feel free to support us. we aim at doing more.

and yes, i would say that @elonmusk mission with @SpaceXAI is going well. the results definitively shows how Grok 4.5 training pushed even further to become as neutral as possible and it does feel good to know there is at least 1 model out there that is within that neutral zone.