Tanmay Patil

🚀 Just released: Krea-2 Depth ControlNet-LoRA Keeps near-perfect 3D structure while letting you completely change the image with any prompt. Works great with Krea-2 Turbo too. huggingface.co/Patil/Krea-2-d… Big thanks to @edwixxxx @Shauray7 who helped test and refine it along the way. Thanks @krea_ai for great model 🤗.

FA4's SingleTileLPTScheduler exploits that causal attention work grows with block index, so it just visits blocks in reverse (block = num_block - 1 - block). So why not try something similar on grouped gemm! In grouped GEMM the analog is that a tile's mainloop time is proportional to its group's K, and StaticPersistentGroupTileScheduler visits tiles in group-metadata order. So LPT = order groups by descending K. Result is 1.74x speedup in grouped gemm, just by sorting the scheduling path.

@Kshitijjkapoor Lmao

not sleeping just in case they revoke fable by the time i wake up

@thsottiaux I am ready with 2 usage resets .

Can't wait to see what people will do with GPT-5.6 Sol Ultra. Stash your hardest prompts somewhere.

Cooking something .. let's see how it goes 🥴

@NielsRogge @aiDotEngineer Hey can we have dark mode on papers with code .. 🤗 BTW nice work

A bit busy this week for @aiDotEngineer, but here are the top 3 trending papers of this week: - "Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent" - "Introducing LongCat-2.0" - "Scalable GANs with Transformers" Explore the papers and evals on Papers with Code!

Multi-GPU kernels are the real test for coding models. Today at @aiDotEngineer, @simran_s_arora shared ParallelKernelBench, an open-source benchmark for evaluating whether LLMs can write fast CUDA kernels for real communication-heavy workloads. Proud to see this work from the Together AI Frontier Performance team.

@Zai_org Nice 🫡

Introducing ZCode, the official development environment for GLM-5.2 - GLM Coding Plan subscribers: now 1.5x usage quota in ZCode - BYOK supported: works with your existing subscriptions and APIs - Available on macOS, Windows, and Linux Download now: zcode.z.ai/en

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

@fraserpricee @AnthropicAI time to cancel subscription

My car isn't getting washed up but @AnthropicAI certainly is 🫵😂 Top is DeepSeek-V4-Flash-DSpark running @ 250TPS on 2 local GPUs Bottom is Sonnet 5, which tbf should definitely be regulated because it is going to ruin a LOT of codebases

@huggingface Buckets are awesome

Introducing Claude Sonnet 5, our most agentic Sonnet yet. It makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models.

It's super interesting to know the system architecture of the TPUs. henryhmko.github.io/posts/tpu/tpu.…

👀 vLLM community is working non-stop to get @deepseek_ai's new DSpark spec decode algorithm for vLLM! Faster inference for everyone! github.com/vllm-project/v…

Chess engines tell you the best move. But grandmasters are human, they don’t always play it. So I built "Kibitz": a human move predictor for chess broadcasts. I trained this model on my Nvidia RTX 5080. Then I made it run as a business by itself. A channel buys the overlay, Hermes onboards them, charges via @stripe test mode, runs the broadcast, narrates with @NVIDIAAI Nemotron, tracks inference cost, and books its own P&L. I build. Hermes operates. This is my demo and entry for the @NousResearch × @NVIDIAAI × @stripe Hermes Agent Accelerated Business Hackathon.

Type of shit I was doing in 2013 instead of learning CUDA