SWE-bench Verified has become saturated. Seven models scored 95–97% (Vals AI, Sep 2026). But how well can coding agents build complete software projects from scratch? Zero2Repo is a Gradient Data research initiative building hard, verifiable tasks and training data for whole-project development. Agents start with a full spec, an interface contract and an empty workspace. They build the entire repo without reference code, then submit it for hidden deterministic testing. Every test must pass for a task to count. No LLM judge is used. The first release includes 17 tasks derived from real open-source projects across Python, TypeScript, Go, C, C++ and Zig. Across 18 model–harness configurations, scores range from 1/17 to 15/17. The top 10 span 47 percentage points, compared with 5 points on SWE-bench Verified. No configuration has completed all 17 tasks. This matters for self-iteration: small unresolved issues can persist and compound as agents continue developing a project. Reliable self-evolution remains an open research challenge. We’re expanding the task collection and looking for contributors to suggest seed repositories, develop tasks and review requirements and hidden tests. More cases are coming. 👉 zero2repo.ai 👉 github.com/Gradient-Data/Zer…

These 17 tasks are a starting point. We’re inviting developers, researchers and open-source maintainers to help expand Zero2Repo. Recommend a seed repo, propose a task, or review requirements and hidden tests. Contribute: github.com/Gradient-Data/Zer… Details: zero2repo.ai

Each Zero2Repo task starts with a spec, interface contract and empty workspace. Agents build a complete repo from scratch, then submit it for hidden deterministic testing. A task counts only if all tests pass. No LLM judge is used. Across 17 tasks in 6 languages, 16 model–harness runs scored 1/17–15/17.

One area we’re focusing on is software engineering. Seven models scored 95–97% on SWE-bench Verified. (Vals AI, Sep 2026) But can they build full projects from scratch? Zero2Repo, Gradient Data’s latest research initiative, builds verifiable tasks for training and evaluation. zero2repo.ai

Over the past year, we’ve worked on Echo and Logits to make AI post-training more efficient and accessible. Gradient Data now extends that work with verifiable environments, real-world trajectories and evaluation systems for more capable and reliable agents. gradientdata.ai

These projects span different layers of the stack, but they come from the same belief: better AI systems require better infrastructure, better agents, and better ways to measure whether they work. Grateful to everyone who researched, built, tested, revised, and kept pushing alongside us. Can’t wait to see what’s next.

EVM-QuestBench, our ACL 2026 submission. For onchain agents, code that looks correct isn’t enough. The benchmark evaluates natural-language transaction code by executing it on a forked EVM environment and verifying whether the resulting state actually matches the user’s intent. arxiv.org/abs/2601.06565

TwinRouterBench Accepted to the NeurIPS 2026 Evaluations & Datasets Track. As more capable models become available, routing becomes increasingly important. But routing needs better evaluation too. TwinRouterBench benchmarks model selection at the level where agentic systems actually make decisions: each step.

DynaWeb Accepted to EMNLP 2026 Main Conference. Training web agents on the live internet is expensive, slow, and difficult to scale. DynaWeb takes a different route: learn a world model of the web, then let agents improve through simulated interaction.

Echo-2 Accepted to NeurIPS 2026 Main Track. Scaling reinforcement learning isn’t only a modeling problem. It’s an infrastructure problem too. Echo-2 explores distributed rollout generation for more cost-efficient LLM post-training, while preserving training quality.

A few research milestones worth looking back on as September comes to a close. Our team and collaborators have been working across model infrastructure, agents, and evaluation, from making RL more efficient to testing whether AI systems actually do what they intend. 🧵

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 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: arxiv.org/abs/2601.22149 Code: github.com/jadeleiyu/MBRL-Ag…

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. github.com/GradientHQ/sympho…

More to come.

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.