You always think you're safe until your job becomes a benchmark.

Beyond removing refusals, I hope these techniques can be used for "latent fine-tuning" to customize models at inference time. https://t.co/rqacwXnmzf

Abliterate LLMs with Heretic 1.1 It's cool to see this project evolving into a solid open-source library The new viz feature shows how the abliteration process gradually groups the residual vectors into two nice clusters 👀

@asapzzhou More details in @asapzzhou's thread, please give him a like :)

@asapzzhou Here are checkpoints made with this recipe you can actually try in dLLM

Open recipe to turn Qwen3 into a diffusion LLM 👀👀 > Swap the causal mask for bidirectional attention > Source model matters a lot for performance > Block diffusion (BD3LM) >> masked diffusion (MDLM) > Light SFT with masking Great work from @asapzzhou with his dLLM library!

The "commodity AI" thesis is wrong. The API market is splitting into two modalities: - Premium models (Claude) dominate programming and high-stakes work. Users pay $2+/M tokens because correct code > cheap code. - Cheap open models own roleplay and creative tasks. Volume is

Please release the forbidden training dataset.

Proudly powered by LFM2 for the language backbone Compared to Qwen2.5-1.5B, it achieves 2.9x higher throughput and 4x larger context length on NPU hardware

Today we introduce Liquid Labs, our advanced research unit, with the goal of understanding and building efficient and adaptive intelligence systems. Liquid Labs consolidates our existing research efforts at Liquid across architecture of foundation models, multimodality,

You don't understand evals Everybody in AI should read this 👇

https://t.co/Y0OvO3DPZG

LFM2 Technical Report dropped! 🥳 It provides details about the LFM2 architecture, pre-training, post-training, vision, audio, and ColBERT models It's 51 pages long, have fun!

🚨New Blog Alert: Is AdamW an overkill for RLVR? We found that vanilla SGD is 1. As performant as AdamW, 2. 36x more parameter efficient naturally. (much more than a rank 1 lora) 🤯 Looks like a "free lunch". Maybe It’s time to rethink the optimizers for RLVR 🧵

Here's the calculation without FP32 gradient accumulation: - Model parameters (FP32): 4 bytes per param - Gradients (FP32): 4 bytes per param - Adam's optimizer states (momentum + variance): 8 bytes per param - SGD: 0! (Still have to add activations on top.)

SGD without momentum is beautiful

📝 Blog post:

Does SGD > AdamW for RLVR? > RLVR updates very few parameters (sparse subnetwork) > The "active" parameters may share similar properties > Similar loss curvature → single learning rate sufficient > SGD (uniform LR) ≈ AdamW (adaptive LR) It means RLFT on potato GPUs!

@nvidia 📝 Paper: https://t.co/Ub8MkEdp8Z (Please ignore the benchmaxx, it doesn't matter in this case)

ToolOrchestra is such a cool work from @nvidia Just an 8B model trained on calling tools and other LLMs to answer queries It's a great demo of what frontier SLMs will be about in 2026