🚨 WHAT THE FUCK IS ALIBABA DOING TO QWEN 🚨 A former Alibaba Cloud employee just revealed what sounds like an absolute shitshow behind the scenes. They have the GPUs, the talent and one of the best open weight model families in the world. And what do the geniuses upstairs do? Keep replacing executives, restructuring entire teams, pushing out random AI apps and worrying about token revenue before the fucking models are even ready. HOW IS THIS REAL 😭 I’ve been hyping Qwen 4 for TWO FUCKING WEEKS. It looks insanely promising, it’s supposed to be dirt cheap AND open weight. Literally everything I want from a model. I genuinely thought Alibaba was about to cook something huge. AND NOW I’M READING THIS SHIT??? These mfs have everything they need to win and somehow the people running the company might be their biggest fucking problem. QWEN 4 ISN’T EVEN OUT YET AND ALIBABA IS ALREADY SPEEDRUNNING THE MANAGEMENT DISASTER 💀 JUST LET THE FUCKING TEAM COOK.

We've given all Pro, Max, Ultra, and Team users one free full reset, across all models and limits. Use it anytime before December 10 at 2:00PM PT. We recently fixed a few small issues that used more limits than intended, and the Fable resets didn't do much for those of you who'd already hit your weekly limits for all models. While these issues didn't affect everyone, we want to make up for the trouble. You'll find it in Settings > Usage, or at

Anthropic says its Claude Mythos model got Claude processing prompts on AMD's MI355X GPUs in about 3 days, with almost no human help. That job used to take dozens of engineers 6 to 12 months. Claude wrote or adapted almost all the specialized GPU code needed to make it work.

DeepSeek V4.1 Flash 1017 on one Mac Studio M5 Ultra 512 GB, official lossless weights: 65.7 tok/s output with a 128k-token prompt, mlx-serve 26.11.2 defaults. Median of 5 runs, range 62.2–70.0. Other engines and the method below.

Drafter speedups are not a 256 token demo trick. M5 Max, TensorFold, Qwen3.8 27B: 129 tok/s across 162K tokens of long output, against 24 tok/s with drafting off. 30 runs, 2.7K to 8.2K tokens each, three sampling settings. 5x holds when the answers get long.

SoftBank defaults on power turbines and electrical equipment for Ohio data center, resulting in $3.7 billion loss on the orders. It’s believed remaining $6.5 billion in order was for $NVDA GPUs that they will attempt to procure to $ORCL on contract. No specific details given. SoftBank promises though it won’t reduce the size of the Ohio data center that never even made it to permit stage before default. OpenAi $CRWV $IREN bankrupt 2027

DSpine: +27.8% acceptance length and +23.3% throughput over DFlash on Qwen3-8B, according to the paper. But the gain is much smaller with the released checkpoint. It was trained on Open-PerfectBlend, and DFlash2 and DSpark retrained on the same data are close behind: about 4% lower at temperature 0 and 5% lower at temperature 1. Still a real gain, for a round that costs about the same as DFlash2's. If you already serve Qwen3-8B or Qwen3.8-27B with SGLang, it's worth an A/B test. I would measure tokens/s on my own prompts rather than trust acceptance length on GSM8K...

Moonshot's open-weight Kimi K3.1 Pro just pulled off a clean sweep on Artificial Analysis's enterprise capability rankings. Across Finance, Strategy, Legal, Healthcare, Engineering, and Economics, Kimi K3.1 Pro ranked #1# in every single category among open models beating Ling 3.1 Flash, DeepSeek V4.1 Flash 1017, GLM-5.3, and Meta's Muse Spark Open. Plus, SGLang merged PR #49472# today adding "breakable CUDA graphs" specifically for Kimi K3.1's hybrid linear attention, cutting prefill latency in half (-53%).

Anthropic put GPU kernel optimization in the same Safeguards section as hacking and biology. WTF? It is in their launch post. During training they limited the skills that "could most directly accelerate AI development". "Accelerator kernel optimization." "Improving large language model architectures." There Fable 5.2 stays "at about the level of Opus 5.5". So making a model run faster on a cheap GPU is now a risk to be managed. You cannot see this one. There is no refusal and no warning. It was limited during training, before any user typed a word. Kernel work is how a used 3090 still runs new models. It is how open weights get to people who cannot pay $40 per 1M output tokens. What the hell? I genuinely wanted to say something nice, but now I can't. This completely broke me, I'm disappointed.