Qwen 3.8 Max: An open model just beat GPT and Claude at using a computer.
OSWorld tests if an AI can click buttons and finish real desktop tasks like a person.
Ok skalskip92 didn't lie, Qwen 3.8 Max has impressive vision capabilities!
The few-shot image prompting is amazing: you can annotate an image with one bounding box, and it will find the remaining objects 🤯 I tried it out on…
Qwen 3.8: Alibaba just dropped a 2.4 trillion parameter AI.
And you can run it yourself. It's one of the biggest models ever released. Open. Not locked behind one company.
Qwen 3.8 Max Full COURSE 1 HOUR (Build & Automate Anything)
However 3d modeling Qwen 27b BF16 just nails it.
Muse puts up a good show but it's not really close here. However like always Qwen uses way too many tokens in general, I wonder how 3.8 is going to…
This was made with MiniMax_AI M3 model i was testing the harness and thought i should try the boss…
This was made with MiniMax_AI M3 model i was testing the harness and thought i should try the boss fight benchmark and I'm really impressed, this is…
Anything under deepseek v4 flash basically doesn't exist to me.
With <200GB of vram, i'm getting ~800K context @ 200 tokens per second. It's not quite Cerebras fast, but damn.
Me waiting for Qwen 3.8 27b...
Meet Qwen 3.8-Max..
Your always -on workmate for: - Coding - Reasoning - Researching - Writing - AI Agent With Qwen 3.8-Max: Just live your life to the fullest.
A very basic front-end web test for Meta Muse Glimmer!
From memory, this result is probably a little better than stock Qwen 3.5 out of the box. It's not perfect, but it's pleasant!
Muse Glimmer just dropped and I got it running on dual 3090s with NV Link right now, stacked with the D flash on top.
🔥 Without D flash it was pushing 40-60 tokens/sec. With it? We're hitting 160-180 on the peaks before it settles.
First impressions on Muse Glimmer!
It's incredibly fast for a dense model, currently running an average of 208tps with a max of 274tps on a single 5090 with their DFLASH config.
Um modelo Qwen misterioso chamado Kiana apareceu no Arena e os testes iniciais colocam ele no mesmo patamar do Opus 5 e GPT-5.6 Sol.
Nome estranho, benchmark forte. Alguém já testou isso?
Qwen 3.8 Max is huge
2.4T total parameters. 95B active during inference. 1M token context. But parameter count isn't the important number.
Qwen 3.8 Max: A free open model just beat GPT and Claude at using a computer.
The test: click buttons, open apps, finish tasks on a real desktop. → Qwen 3.8 Max scored 86.1 → GPT 5.6 scored 83.2. Gemini scored 76.2.
Okay. Wildddddd!! Grok 4.5 beat Kimi K3 on the exact same task for 13x cheaper 👀
Aimlapi ran the test: same prompt, same one-shot job, no tricks Cost per figure: Grok 4.5 → $0.15 GPT 5.6 Sol → $0.60 Qwen 3.8 Max → $0.67 Kimi K3 →…
Qwen 3.8: There's a free 2.4 trillion parameter AI that goes head-to-head with the biggest names.
It just landed, and most people have no idea it exists. → Reads text, images, and video → Holds 1M tokens of context → Works on projects for days…
Qwen 3.8 Max Full COURSE 1 HOUR (Build & Automate Anything)
The biggest upgrade isn't another model.
It's shared memory. Every Hermes conversation gets written into an Obsidian vault.
Zuckerberg is finally building AI.
China ran out of GPUs. Microsoft is losing relevance. The American AI monopoly just ended.