You always think you're safe until your job becomes a benchmark.
@maximelabonne
-
Latent Fine-Tuning: Customizing Models at Inference Time
By
–
Beyond removing refusals, I hope these techniques can be used for "latent fine-tuning" to customize models at inference time.
-
Heretic 1.1: Open-Source LLM Abliteration Library Evolution
By
–
Abliterate LLMs with Heretic 1.1
— Maxime Labonne @ ICLR (@maximelabonne) 11 décembre 2025
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 👀 pic.twitter.com/PLYmRPuYB7Abliterate 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
-
Converting Qwen3 to Diffusion LLM: Block Diffusion Method
By
–
Open recipe to turn Qwen3 into a diffusion LLM 👀👀
— Maxime Labonne @ ICLR (@maximelabonne) 8 décembre 2025
> 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! pic.twitter.com/ec2tSXAUA1Open 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! -
DeepSeek Model Distillation Continues Strong Momentum
By
–
Looks like the DeepSeek distillation is still going strong
-
OpenRouter Token Usage Analysis Roleplay Breakdown
By
–
They've always used OpenRouter, as shown by the roleplay breakdown graph (it's always been higher than 20% of total tokens)
-
AI Market Splits Premium Models Versus Cheap Open Source
By
–
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
-
6D Parallelism: The Human-like Parallel Processing in AI
By
–
Little-known fact in AI: 6D parallelism is human parallelism
-
TPE Optimization Performs Well for Small Medium Models
By
–
I tested and in practice it's really okay for small to medium models. I'm also a bit biased because I'm a big fan of TPE from my PhD days.
-
Shopify Scales LFMs: Untapped Potential of Small Models
By
–
Super proud that @Shopify now uses LFMs at scale! The potential of small models is still largely untapped, we're just getting started