There’s a good chance your open source model is costing more than the frontier. Cheap tokens ≠ cheap tasks. Here, we introduce Intelligence Density, and Density Aware Training, our post-training technique to achieve less wasted compute, better learning, all with no knobs to tune. Enabled by default in every Trajectory model.

At Trajectory, we're constantly implementing and building upon the latest research ideas on the path to continual learning. We wish we had the time to share all of them, but here's a quick glimpse on our explorations with PiSSA, and choosing the right trainable geometries for agentic RL.

Really nice post on how to decide what to work on, embodies many of the reasons I decided to work at @trajectorylabs and why we’re proud to partner with @thinkymachines on building better open source models. 10/10 name collision

At Trajectory, we care about storytelling. The storytelling about continual learning, the storytelling about the research breakthroughs it’ll take to get there, and the storytelling about the product that we need to will into existence. Brand is part of how we tell it. Here’s a behind-the-scenes look at the work we did with @metalab to craft ours

The bitter lesson only applies to methods, not to verifiers. Distilling ambiguity into quantified directional signal is the first step in solving real world problems with computation. In this talk at Weights Unknown, I describe the annoyingly simple eval framework we use at Harvey to build P-sets for legal expertise, and how we lean on the frontier ecosystem to find the methods that optimize against them. Credit for most of the work behind the talk goes to @nikogrupen, @ItsJulioPereyra, @gabepereyra, and frontier ecosystem partners @appliedcompute, @baseten, @EngramLab, @FireworksAI_HQ, @trajectorylabs (+ many more!) Thanks @AmplifyPartners for having me!

Had a great time speaking at the SkyRL event on how Trajectory built #TCLI# to serve hundreds of RL training runs a day, so our users' models keep getting better. - Shout out to our partners at @AMD and @wafer_ai for sharing the exciting disaggregated RL result on climbing AIME benchmarks by rolling out on AMD and training on NVIDIA. - And as always good to hear more deep dives on how SkyRL works underneath the hood (multi-LoRA, harbor integration) from @erictang000 @sumanthrh and @charlie_ruan Btw if anyone can guess what TCLI stands for I'll buy them a burrito. 🌯 Hint: there are two right answers.

It's time to rethink RL. Translating real world use into model improvements requires redesigning post-training algorithms for non-verifiable, per token rewards. At @aiDotEngineer 's World Fair, we share our insights into scaling algorithms like SDPO for continual learning.

Continual learning is a bet that the retraining loop will get cheaper over time. With larger models, you can maybe run this loop once every few weeks. But with smaller models, you can run it nightly, per customer. And it keeps recursing: a model per company, then a model per client that company serves, then per matter. We’re getting closer to intelligence cheap enough to meter. On the path to this, we received early access to, and post-trained @nvidia's Nemotron 3.5 Lightning on @harvey LAB. One click on the Trajectory platform, no new engineering. 0% to 8.3%, above Opus 4.6 at 6.6%.

We're excited to sign the call for Open Weights. We believe the best way to create something enduring is to start with the future you believe is coming, then work backwards. We think the future is one where every product has its own intelligence, shaped by its users, its workflows, and everything it learns after it’s deployed. We’re building the experience layer for that future, and the products to bring that control into everyone's hands. However, in almost every path we can imagine to that future, open weights play a major role. Not because every model will be open, but because they give builders ownership over one of the most important layers of the stack. The more capable open models become, the more ambitious the products built on top of them can be. We’re excited to do our part to help make that future happen.

1/ We post-trained @nvidia Nemotron 3 Ultra on @harvey Legal Agent Bench in under 24 hours. The result: an open model reaching the same band as leading closed models on legal work, at a fraction of the cost. The correlating story: when a new open model ships, Trajectory can turn it into a specialized agent almost immediately.

5 Days of Trajectory 🏹Day 5: Scaling SDPO to Agentic Tasks Continual learning means you must train on data from production. But production gives you one example per task. A user makes a request once. You get one trajectory, not a batch. However, current RL algorithms don't work that way, They need groups of tasks. By definition, that means you need some artificial environment to perform those rollouts in. But what if you don't? SDPO is a promising route. It learns from a single trajectory, with no group required and failures still producing signal. The shape of the method matches the shape of production data. But one fundamental problem remained. Every published SDPO work assumed fresh, on-policy rollouts. Agentic work cannot give you that. Trajectories run for an hour or more and arrive stale. On true agentic tasks, naive SDPO collapses. We fixed it. We're the first to make SDPO work on agentic tasks. On Mercor's APEX-Agents, with hour-long trajectories and near-zero base pass rates: 25% average reward, 5x over zero-shot. More importantly, it trains stably and the curve is still climbing. Read more below.

🏹 5 Days of Trajectory. Day 4 - Why We’re Building Trajectory AI is the most capable software ever built. You correct it. You teach it what you want. However, the next session starts, and the learning is gone. This is deeply unnatural - nothing intelligent works this way. Today, we’re sharing the thesis behind Trajectory: - why continual learning is the next platform shift in AI - why the primitive governing that shift is the trajectory - our plan to move products from being shipped to being grown: first make the intelligence layer better, faster, and cheaper; then make it shapeable; finally, make it learn Read more below⬇️

We’re taking a quick break for the 5 days of Trajectory, but wanted to take this time to say that we’ve been named to @Redpoint’s 2026 Infrared 100 as one of the companies shaping the future of AI infrastructure. We're so grateful for the recognition so early in our journey, and want to congratulate the other awardees as well!

Welcome to Day 2. Yesterday, we showed the broader work we're doing with the pioneers of continual learning. Today we'd like to deep dive on one: how we post-trained an open model for legal work, in partnership with @Harvey. We've built a platform where production data is the moat. Every correction, retry, and edit becomes signal you can post-train on, and the models are plug and play: customer's can drop in their model of choice, and improve from there. Fields like legal and finance make those demands absolute, with hard security, sovereignty, and provenance requirements. That's why we post-trained @nvidia 's open-weight Nemotron 3 Super, on Harvey's LAB benchmark. The results, in just hours: post-trained Nemotron 3 Super approaches the closed frontier, matches GPT 5.5, lifts rubric-pass criteria +25%, all while beating the performance-vs-cost frontier. That's the power of our platform. And this is just a glimpse towards what the future of intelligence will look like: continual learning, where products get smarter every time they're used. Thanks to @nikogrupen, @gabepereyra, @ItsJulioPereyra, and the whole Harvey team for their collaboration on this. Much more to come soon on continually learning legal agents