Sapient Intelligence

Building leaner architectures for deeper intelligence.

[Sapient Intelligence @Sapient_Int]

Introducing PRAXIST Beta, your autonomous research team🚀

Define the objective, constraints, and what success looks like. PRAXIST discovers the path.

From there, PRAXIST takes on the experimental research loop. Multiple Research Peers explore competing approaches in parallel, share useful findings, and build on accumulated evidence across experiments and generations.

In partner-provided environments, PRAXIST reached a 100% safe-landing rate in a rocket simulation and reduced accumulated error in an industrial SLAM system from 9.37 centimeters to 5.01 centimeters.

From robotics control to quantitative finance, PRAXIST has delivered measurable advances across fundamentally different problems by combining a shared core research architecture with domain-specific tools, knowledge, constraints, and evaluators.

More approaches explored. Faster iteration. Greater R&D capacity without proportionally increasing specialist headcount.

[Sapient Intelligence @Sapient_Int]

We’re happy to share that two papers from Sapient Intelligence have been accepted to NeurIPS 2026!

The accepted work contributes to a broader research agenda around optimization, adaptation, and efficient learning in complex systems, all of which are key to building increasingly capable and autonomous learning systems.

See you at NeurIPS 🎉

[Sapient Intelligence @Sapient_Int]

Curious to try recurrent reasoning? Get started with HRM-Text, with open-source code, model weights, and a full pretraining framework 👇

Hugging Face: https://t.co/X7DW812tq2
GitHub: https://t.co/GKR8vFJZND

[Sapient Intelligence @Sapient_Int]

Excited to see the conversation around OpenAI’s Astra bringing recurrent architectures into the frontier spotlight.

At Sapient, recurrence has been central to our vision from day one. We introduced HRM in June 2025 and open-sourced HRM-Text this May. It’s encouraging to see a direction we committed to early gaining wider attention.

Want to explore recurrent reasoning yourself? HRM-Text is an open-source ~1B-parameter model trained from scratch on just 40B unique tokens, for about $1,000 in GPU compute-with a full pretraining framework to get you started.

Explore the architecture. Train a model. Build on it.

There’s still so much to explore in recurrent architectures. We’re continuing that work and building PRAXIST to help more researchers pursue ideas they believe in, just as we have with HRM.

[Sapient Intelligence @Sapient_Int]

Have feedback, questions, or ideas for PRAXIST? Join our Discord community to discuss directly with the team and other builders.

We’d love to hear what you’re working on—and what you’d like to see next.

Discord: https://t.co/sW0boIQPl7

[Sapient Intelligence @Sapient_Int]

We’re still learning how to maintain and evolve PRAXIST responsibly, and we’ve established the following contribution workflow:

1. Submit an issue or PR
2. Maintainers review it, and proposed fixes move into validation branches
3. Related PRs may be grouped into a focused validation branch
4. Each branch goes through CI and real-world task testing, typically requiring 1–3 days of validation by the PRAXIST maintenance team
5. Once validated, changes are merged into main, validation branches are removed, and contributors are credited

We aim to run this cycle on a weekly basis. We’re grateful for every piece of feedback, idea, and contribution that helps make PRAXIST better.

More updates are coming soon. Stay tuned.

🔗 Learn more about the contribution workflow:
https://t.co/0N7PutMgGQ

[Sapient Intelligence @Sapient_Int]

Complex engineering solutions are rarely the result of a single breakthrough. They are assembled from discoveries: a mechanism uncovered in one experiment, a design principle validated in another, or an interaction that later becomes useful in a completely different direction. PRAXIST preserves these discoveries as evidence-backed building blocks that carry across generations of research, where they can be validated, refined, reused, and combined into more capable designs.

With every generation, the set of tested building blocks grows, giving the next round of research a stronger foundation to build on. Useful discoveries can be carried forward, refined, and recombined across future experiments, so each generation begins with more knowledge than the last. Every experiment expands what can be built next.

Read the full paper: https://t.co/QBoybSSIkI

[Sapient Intelligence @Sapient_Int]

From the stage to the booth, PRAXIST sparked a lot of conversations at AI4.

We spent the week meeting leaders and technical teams from across industries, discussing the problems they’re working on and how PRAXIST could be applied inside their organizations.

Thanks to everyone who stopped by, shared their challenges with us, and explored what putting autonomous R&D to work could look like.

[Sapient Intelligence @Sapient_Int]

PRAXIST on stage at Ai4.

Earlier this month, our co-founder @wcwilliamchen shared the thinking behind PRAXIST and our approach to the next generation of autonomous R&D.

A few moments from the presentation.

[Sapient Intelligence @Sapient_Int]

What does it look like when AI learns to land a rocket?

PRAXIST coordinates a team of Research Peers, each taking on a different part of the problem. Together, they learn across experiments and turn accumulated evidence into a validated, production-ready solution.

See it in action🚀

[Sapient Intelligence @Sapient_Int]

@jumperz The PI Panel looks beyond the current best result. If a direction shows a meaningful early signal but the evidence is still weak or noisy, it may run another experiment to validate, falsify, or better characterize the finding before deciding whether to pursue it further.

[Sapient Intelligence @Sapient_Int]

Introducing PRAXIST Beta, your autonomous research team🚀

Define the objective, constraints, and what success looks like. PRAXIST discovers the path.

From there, PRAXIST takes on the experimental research loop. Multiple Research Peers explore competing approaches in parallel, share useful findings, and build on accumulated evidence across experiments and generations.

In partner-provided environments, PRAXIST reached a 100% safe-landing rate in a rocket simulation and reduced accumulated error in an industrial SLAM system from 9.37 centimeters to 5.01 centimeters.

From robotics control to quantitative finance, PRAXIST has delivered measurable advances across fundamentally different problems by combining a shared core research architecture with domain-specific tools, knowledge, constraints, and evaluators.

More approaches explored. Faster iteration. Greater R&D capacity without proportionally increasing specialist headcount.

[Sapient Intelligence @Sapient_Int]

The PI Panel doesn’t force an early consensus. When parallel peers produce conflicting signals in a high-noise space, it treats the evidence as unresolved and designs more targeted follow-up experiments to isolate the effects, reduce confounds, and determine which findings hold up.

[Sapient Intelligence @Sapient_Int]

Today, evaluators and constraints are still largely domain-defined. Domain experts specify what matters, what should be tested, and what counts as a meaningful result. PRAXIST can then explore within that framework, but it does not yet generate the full evaluator stack from the objective alone.

[Sapient Intelligence @Sapient_Int]

@exploraX_ Shared memory carries findings rather than raw experimental outputs. PRAXIST captures what was learned, including relevant context around why a direction worked or failed, so other peers can build on the useful evidence without inheriting all of the experimental noise.

[Sapient Intelligence @Sapient_Int]

From robotics control to quantitative finance, PRAXIST has delivered measurable advances across fundamentally different problems by combining a shared core research architecture with domain-specific tools, knowledge, constraints, and evaluators.

Across more than 100 research and optimization tasks, PRAXIST has been applied to problems spanning over 50 application areas.

[Sapient Intelligence @Sapient_Int]

Across the full 75-task MLE-bench, PRAXIST (with an open-source model) achieved:

→ 65.3% Gold Rate — 49 gold medals
→ 80% Any Medal Rate — 60 total medals
→ 44% more gold medals than Claude Code + Claude Opus 4.8
→ $3,054 in token cost, compared with $38,370 for Claude Code + Claude Opus 4.8 — roughly one-twelfth the cost

Against the stronger Claude Code baseline, PRAXIST delivered better benchmark results at 92% lower token cost.

*The token cost in USD for the full 75-task benchmark suite, measured July 1–30, 2026.

[Sapient Intelligence @Sapient_Int]

General-purpose coding agents are optimized to complete software tasks. Many research agents use a winner-takes-all loop: keep the best result and discard the rest.

PRAXIST is purpose-built for cumulative experimental research. Every useful finding can shape what happens next.

[Sapient Intelligence @Sapient_Int]

PRAXIST assembles the team each problem needs: analyzers, explorers, innovators, exploiters, falsifiers, and combinators.

Research Peers pursue competing directions in parallel. Shared memory connects their evidence. A PI Panel decides what to test next.

[Sapient Intelligence @Sapient_Int]

Starting today, PRAXIST is open source.
Explore the code. Adapt it to your domain. Discover what comes next.

Learn More: https://t.co/pWHBxG9Tnd
GitHub: https://t.co/RanTG5UCOU
Tech Report: https://t.co/LZ75WjhE6T
Announcement: https://t.co/drhwlVtOpo

[Sapient Intelligence @Sapient_Int]

LLMs generate.

Agents execute.

But what happens when the path itself is unknown?

It must be discovered.

Explore. Experiment. Validate. Optimize.

Stay tuned.