thanks to everyone who joined my session at @aiDotEngineer paris yesterday! for everyone who missed it, here's the thread form of my talk 🧵

MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today. Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once. Compared with MiMo-V2.6's Hybrid SWA architecture: • 5.02× lower prefill FLOPs at 1M tokens • 4.5× smaller KV cache at 1M tokens • Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL Why build a new architecture? Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-cache size, and retrieval accuracy on the critical path at the same time. HySparse2 tackles all three with two levels of KV sharing: • KV Bridging: Following YOCO, full-attention layers in the cross-decoder build their K/V from self-decoder hidden states. • KV Reuse: Within each hybrid block, sparse layers reuse the preceding full-attention layer's KV cache and selection indices. Two more changes: token-level selection replaces block-level selection, and a forced window of recent tokens replaces the separate SWA branch, so local and global tokens share one KV cache. Since all cross-decoder KV caches now come from the self-decoder, prefill can stop once the self-decoder finishes. Paper: arxiv.org/pdf/2609.26368

Who’s responsible for rebranding classifiers as “decision models” ??

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 🧵

We've reduced token costs in Cursor by 7% with no drop in agent quality. Savings came from tighter prompts, selective tool loading, better caching, and compressed file reads.

We’re demonstrating how frontier models have continued to improve in realistic mental health conversations with MentalHealthBench. This new open benchmark was built with input from more than 80 mental health clinicians. We’re releasing it openly so other researchers can examine the methods, run their own evaluations, and build on the work. openai.com/index/introducing…

🚰 SYSTEM PROMPT LEAK 🚰 Here's the full system prompt for Claude Opus-5.5!! The total count of everything extracted, including all tools, comes in at over 1.9M characters! 🤯 Lots to dig into here. Enjoy! 🫡 Link: github.com/elder-plinius/CL4… gg

The recent breakthrough in Navier-Stokes has garnered a lot of attention to the new possibilities that emerge when agents work together. We have been interested in this question for a while. How can a team of agents achieve more than agents working alone? I'm excited to finally share our new work, "Self-Organizing Agent Teams Learn to Reason Together" 🧵

Here's what the historical open vs closed data here looks like. Would be useful to have volume too (not just %)!

There's a fun mix of excitement and cope on my timeline re-Jev: - OMG I made Jev a year ago this is dumb - You can fine tune this encoder into a better, faster, local classifier, etc Re the "just train your own small classifier" crew - Voyage trains great, SOTA pointwise rerankers that are obviously specialized to the reranking task and have been at it for years, likely also using a strong synthetic data recipe... Exact setup in the response tweet. Do note that using Jev is definitely more expensive than using Voyage rerank-3 bc of the tokens in the lengthy rubric included in every request (though you could definitely optimize this further) Jev is interesting to me bc I didn't expect it to be able to outperform domain specific models like these rerankers. Every other OpenJev etc type model I've benchmarked gets ~15% agreement at 20 on this eval. It's almost like it would be tough to perfect the data recipe to train a general classifier 🤔

Smaller subset of 200 queries (prior post was >1k) but jina-reranker-v3.5 doesn't do well here and more notably, voyage rerank-3 is a good baseline bc it's ~Luna quality (Luna (low) with blitzrank is wayyyy higher latency to be clear, this is a method for prepping a dataset not a reranker you'd use on prod). If anything Luna should have an advantage here bc it's likely distilled from/trained on much of the same data as Sol's preferences which is my ground truth here so in that sense your assumption should be that this eval makes Luna look better that it should. So here again, Jev is outperforming Luna at lower cost and much much lower latency.