Bismuth
83 posts
Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params. Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale. Model weights: huggingface.co/moonshotai/Kim… Tech report: github.com/MoonshotAI/Kim… Tech blog: kimi.com/blog/kimi-k3
I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!
The coalition of open AI. Let's stand up for Open Source AI!
Andrej Karpathy just said your kids should ignore most of school. "80% of education should be math, physics, CS." Not because it's useful - because it carves grooves in the brain. Grooves that get harder to carve the older you get. In a pre-AGI world: it gets you a job. In a post-AGI world: it makes you a functioning, empowered human. Everything else? Tack it on later. The cognitive foundation is the only foundation that survives the transition.
@pli_cachete No, he should apologize for saying it will pursue its own goals, which it clearly was not (it was doing what the instructions said to do.) If it is that smart and not pursuing its own goals, why should anyone believe making it smarter will make it do that? (No one should.)
🚨 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?
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-…
Gemini 3.6 Flash is here, and it's the best price-to-performance model we've tested for browser agents. On our BU Benchmark it hits 68%, beating GPT-5.6-sol (67%) and Sonnet 4.6 (62%). Second only to Opus 4.8 (74%), at a fraction of the cost. Frontier-level web agents at Flash pricing. @GoogleDeepMind
Gemini 3.6 Flash benchmarks are out, and it's... beaten by other models on code tasks, and is only really consistently SoTA on vision and context benchmarks. But hey, 3.1 Pro is now so old 3.6 Flash outperforms it across the board 😭