ALERT🚨: NVIDIA SGLANG INFERENCE RUBIN IS 4.23x BETTER 🚀 PERF PER DOLLAR THAN EVEN GB300 NVL72. This is on the widely used production LLM engine called @sgl_project. (1/2)🧵

We want people to have the freedom to move to a better job and help our most promising businesses grow. Today, the PM announced we will legislate against non-compete clauses, so they no longer hold back workers or stand in the way of startups hiring the talent they need. We’ll also bring forward options to tackle wider restrictive practices, like excessive notice periods. We have to empower the people taking risks to build something, giving them the freedom to start and scale.

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: contrastive-lm.notion.site 💻 Code: github.com/Contrastive-LM/CL… 🗣️ Discord: discord.gg/5dAQEDJBs 🤗 Data & Models: huggingface.co/Contrastive-L… More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. Streaming the run: mimo.xiaomi.com/rl/

JUST IN: Anthropic refused to give Mythos 5.1 testing access to the UK AI Security Institute (AISI) prior to its release last week, per the FT.

Your gaming PC can now serve frontier models at interactive speed using official checkpoints without extreme quantization! Qwen3.6 35B → 8GB RTX 4060 laptop @ 39 tok/s DeepSeek-V4-Flash 284B → RTX 5090 desktop @ 22-25 tok/s GLM-5.2 753B → RTX PRO 6000 workstation @ 15 tok/s Run your claude code or codex now with frontier model for $0 Meet FreeToken 🧵

Finally wrapped up my work at @AdobeResearch, and excited to share FLARE: Diffusion for Hybrid Language Model🔥 Hybrid attention is having a moment. Recent releases such as Qwen3.8 and Kimi K3 reinforce a compelling direction for efficient LLMs: combine linear attention with periodic full attention. Bounded recurrent states make long-context processing economical, while full-attention layers retain direct global interaction. With FLARE, we asked whether this increasingly important architecture could also support diffusion-style multi-token generation—without rebuilding the model from scratch or giving up the capabilities of its pretrained checkpoint. FLARE transfers Qwen3.5 hybrid-attention checkpoints into diffusion language models with a cheap budget of around 10B training tokens. We release FLARE-2B, FLARE-4B, and FLARE-9B. The key is to train two complementary views of each document: a clean causal stream that preserves next-token behavior, and a noisy block-bidirectional stream that learns diffusion denoising. The resulting checkpoint supports two inference paths: > AR-Trust verifies parallel drafts through the clean causal stream > Diffusion-Trust denoises multiple tokens within an active block Making this transfer work revealed a broader lesson: conversion quality depends on more than the diffusion objective. First, transfer data sets the capability ceiling. Under the same source checkpoint and token budget, changing the balance of reasoning traces, mathematics, code, and instruction-following data reshaped which capabilities survived conversion. A data mixture that helped one capability could hurt another; the best overall transfer came from preserving coverage across them. Second, the clean stream is what anchors the converted model to its source behavior. Clean causal context recovered most of the loss from diffusion-only training, and retaining next-token prediction closed much of the remaining gap. Meanwhile, training across partially completed noisy blocks preserved the flexibility needed for Diffusion-Trust decoding. Hybrid diffusion is also a systems problem. During training, small blocks multiply the number of recurrent-state boundaries that would otherwise be materialized in GPU memory. We developed a fused two-stream route that keeps chunk computation Tensor-Core friendly, stores only strided checkpoints in HBM, and reconstructs boundary states in registers. This makes small-block training substantially more practical and brings its efficiency closer to conventional causal training. Inference requires a different form of state discipline. Paged KV cache and recurrent GDN state must remain synchronized while tokens are still provisional. Our SGLang-based stack commits state at the accepted offset for AR-Trust, keeps denoising rounds read-only for Diffusion-Trust, and updates the causal state only after a block becomes final. Across reasoning, knowledge, instruction following, and coding evaluations, FLARE models preserve a strong capability profile while enabling diffusion-style multi-token generation. With the optimized serving stack, FLARE reaches up to 4.8× higher decoding throughput in our evaluation setup. Everything is open now: the training code, the inference stack, and the FLARE checkpoints. We hope FLARE makes it easier to explore diffusion-style generation on modern hybrid backbones. Try the models, reproduce the results, or use the code to build fast language models of your own. We’re excited to see where the community takes it. This was joint work with @ChongjianG30781, @HaoTan5, @Xu_Arthas, @ZhuWanrong, Jason Kuen, @koustavagoswami, @YongxinChen1, @MoleiTaoMath, and more importantly, with support from my beloved mentors Jiuxiang @ethan_jxgu and Jing @Jingorz at Adobe Research. 📄 Paper: arxiv.org/abs/2606.01774 🌐 Website: tokflare.github.io 💻 Training & inference: github.com/yuchen-zhu-zyc/FL… 🤗 Checkpoints: huggingface.co/collections/y…