And!! I just remembered a few months ago when Elon predicted Chinese labs would get Fable 5 level early next yr & @jietang calmly replied “won’t take that long” Well - it’s Oct and now we have Fable 5.1 level on LOCAL MODEL 💥 Just think about that. Amazing stuff 🧠🚀

💥Kimi is SO BACK!! and - 3.1 is a strong reminder that the interesting scaling story in 2026 is no longer just pre-training FLOPs. This is the same 2.8T / 104B-active MoE as K3 (896 routed experts, 16 active + 2 shared, 69 KDA + 24 Gated MLA layers, AttnRes, Stable LatentMoE, native MXFP4), but the gains come from substantially larger post-training compute and large-scale joint RL across hundreds of domains. The result is a model that can write/iteratively debug its own Blender/CAD scripts, assemble full game projects in Unreal/Unity/Godot, and produce linked financial models with meaningfully lower hallucination rates. On Terminal-Bench 4.0 it sits at 65.7% (between GPT-6 Astra & Opus 5.5), & it leads several long-horizon agent + coding metrics. 🚀 The open-weight date (Nov 4) is welcome, but the practical constraint is still the same: even with extreme sparsity + MXFP4, this is a supernode-class model. The real test will be how well the post-training stack transfers when the community starts fine-tuning and running it under its own constraints. All amazing stuff 🚀

Published today after dug a bit deeper in DeepInfer maths, in case you missed it. These guys at DeepSeek are way, way too smart. 🤯🤯🤯 I don’t know why I didn’t see this but the revelation changes how I see “the race” completely 🧠 5 min read 👇

Been busy all day & didn’t realised there has been a Turks vs Chinese war here. Not sure what it’s abt but I just know countless Chinese have been waiting for all their life to make their (very tough) compulsory history + geography education useful. Today is the day 😂

DeepInfer math just made me realise closely trailing SOTA is a super smart feature, not a bug. FUCKKKK - these guys are way too smart!!! Had to write it all down - a short article - enjoy!

A reminder of what the whale 🐋 has made open: > weights, under MIT > FlashMLA, attention kernels at 580 TFLOPS on H800 > DeepEP, the communication library for MoE > DeepGEMM, FP8 matrix multiply at 1,350+ TFLOPS. > 3FS, an entire distributed file system they built for themselves. > DeepInfer, the in house inference framework and the engine behind V4.1 Flash > deepseek-recipe so your deployment behaves like their API so you can also gain insane margin This is freakin every layer of an AI lab. Model, kernels, communication, storage, serving. Except their training data - there is nothing else really they haven’t made it open to the world. 🌎 I can’t love and thank them enough. 💙🐳

I have many reasons to love how the whale 🐋 is pushing efficiency. One of the top one is that I’ve been building highly efficient digital hard infrastructure. I’ve spent 18 months running around telling people to be efficient - nobody cared. They just want to build on speed cos all excessive costs /inefficiency can be transferred to the customers & the capital market (private & public) was asking abt the volume, not PUE, not long term ROI (and why), not efficiency. Now everybody has to - what we’ve built: 🚀💪

82% margin at rental GPU is insane for any lab - but DeepSeek will own most of the initial supply of Ascends. Achieving this while being 20X cheaper than OAI & A is absolutely insane. And they OSS everything. And OAI & A’s IPO - the public market will shred them in pieces.