More capable models require more routing decisions, and better ways to benchmark how routers execute. That’s what TwinRouterBench is built for. Congrats to the team on the NeurIPS 2026 acceptance.

Can a coding agent choose the right model at every step? Excited to share that our collaboration with @Gradient_HQ, TwinRouterBench, is accepted by NeurIPS 2026! 🎉 We route model calls at each agent step based on what the task needs next. On 100 held-out SWE-bench Verified tasks, our router cut API cost by 53.1% vs. always using Claude-Opus 4.6, while resolving 75 vs. 74 tasks. Paper:  Code: 

Can web agents learn through imagination? DynaWeb: Model-Based Reinforcement Learning of Web Agents has been accepted to EMNLP 2026 Main Conference. Training web agents with online RL means costly, risky interaction with the live web. DynaWeb learns a Web World Model instead, letting agents dream web interactions and train on imagined rollouts. - Learns naturalistic web transitions from agent actions - Generates imagined trajectories in a synthetic web environment - Interleaves imagined rollouts with real expert trajectories - Consistent gains for open-source agents on WebArena and WebVoyager Imagination as a scalable path toward online agentic RL. This is part of our ongoing work at Gradient on distributed and efficient RL: training capable agents without the cost, latency, and risk of centralizing them. Paper: Code:

Proud to share MERA: joint work from a group of brilliant researchers including our very own. It evolves small models with SkillBook + routing + adapters so they take on more of the agent workload over time. Paper + code 👇

Agents are impressive. Multi-agent systems are where things get interesting. As AI systems evolve from single agents into networks of specialists, a new question follows: How do agents decide who does what? That’s a problem we’ve been exploring with Symphony. Symphony brought multi-agent collaboration across heterogeneous environments. To take that further, we released Symphony-Coord, where coordination and specialization adapt through task context, performance, and feedback.

This is why we built Parallax @tryParallax As open models get stronger and agents get more capable, local-first AI becomes much more than a privacy story. It becomes a new way to build with open intelligence that stays close to your data, your tools, and your machines.

A self-evolving agent + a 428B model + 3 Macs = ? Your own AI lab. We ran @MiniMax_AI M3 locally with @tryParallax, right on our desk. Then @GA_agent_ai took over to create a 5-stock portfolio and write it to disk. No cloud. No API bills. Nothing left the machine. Wild to see a ~3K-line agent drive all this with a 400B+ model on local hardware. Thanks to the GenericAgent and MiniMax teams for making local AI feel real.

nvidia going all in on local ai. here's our take: it shouldn't depend on which chip you bought. sparks, macs, the 5090 already on your desk, we cluster across all of it and split your favorite model pipeline-parallel so it runs fully private and local.

We're hiring at Gradient. Building open-source environment infrastructure for our distributed RL training stack — reproducible, scalable to thousand-GPU runs Looking for 1–2 RL Environments engineers / tech leads: You've designed verifiers, built sandboxes for agentic RL rollouts, or shipped RL training data pipelines that survived contact with real training. Domain depth in math, code, agent, tool, or GUI is a plus. PhD not required. Also hiring research interns: PhD / Masters students with hands-on RLHF / RLVR / GRPO / DPO / agentic RL experience. Open-source footprint matters more than paper count. Most intern roles convert post-grad. No age cap. Founding-team-level equity for the right people. DMs open.

Thrilled to see @tryParallax live in production on @Theta_Network. This is exactly why @Gradient_HQ built Parallax: turning the world’s GPU mesh into a sovereign, distributed token factory. Congrats on the milestone! 🫡

When you scale parallel agents, prompt updates degrade fast. The more trajectories you process concurrently, the more generic your learned prompts become. Our researchers worked with @lihanc02 and team on Combee to rethink how aggregation works at scale. Results held up across GEPA and ACE even past 80 concurrent agents. Read more on this research👇

If software no longer needs you to operate it, what does an “application” even mean? That’s what we’re digging into at The Agentic Shift with panels, demos, and speakers from Google, PixVerse, MiniMax + more. SF | Apr 8 Sign up here:

As fellow training nerds, UniPat AI’s approach caught our eyes. Synthesizing future-event data to fix outcome bias and actually beat human prediction markets is defined something worth checking out!

Benchmarks that test what models have memorized are saturating fast. ARC-AGI-3 is asking a harder question: can AI actually learn something new on the fly? One direction we've been exploring: multi-agent orchestration. In our study, coordinating four frontier LLMs across multiple turns consistently matched or outperformed the strongest single model, even on tasks none of them could solve alone. The gap between "best single model" and "best coordination of models" is where a lot of the real progress is hiding. More on our multi-turn, multi-agent orchestration study: