Join us at @PrimeIntellect to build open superintelligence and the infrastructure powering self-improving agents We’re hiring across 25+ roles > Locations: mainly SF, also NYC, London & remote. > Send proof of exceptional ability of something you built > Apply: Roles: Applied Research • Evals & Data • Forward-Deployed x 10 • RL & Agents Research • AI Research Resident • Research Engineer, Distributed Training • Research Engineer, Reinforcement Learning • Research Engineer, RL Infrastructure Compute • Compute Finance & Strategy • Compute Intelligence Engineer • Head of Compute • Solutions Architect, AI Infrastructure • Technical Account Manager, AI Infrastructure Engineering • MTS, Compute Platform • MTS, Full-Stack • MTS, GPU Infrastructure • MTS, Inference • MTS, Sandbox Platform • MTS, Security • MTS, Training Platform Growth • Applied AI: Product Strategy & Revenue • Forward-Deployed AI Strategy • Head of Growth • Head of Marketing Other • Internship • Open Application for Unconventional Talent

Excited to release Prime-Agent A self-improving RLM harness for coding and long-running autonomous tasks. Our goal is to enable everyone to build and use self-improving agents; and the harnesses, models, and infrastructure to support them. Prime-Agent is token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-improving harness state. It combines three ideas we’ve been exploring: >Recursive Language Model–native programmatic tool calling >Persistent multi-agent orchestration >A self-improving continual harness Together, they let the model act on, and improve its own context and harness. @kevinjosethomas @a1zhang @sethkarten and team cooked 🙌 Install it: curl -fsSL | sh

It’s fairly well known frontier labs use value functions now, but we still don't have good open value function infra. Here's something cool from my ongoing internship @MistralAI with @laurence_ai getting value functions to work Introducing - Prime Values > Clean hackable, independent abstractions built on top of prime-rl by @PrimeIntellect > First-class asynchronous value trainer and evaluator nodes, with no trainer bottleneck > Native value warmup support > Streaming replay buffer feeds value model exploiting its greater staleness/reuse tolerance Defaults validated to match or outperform mean-baseline GRPO on both single-turn and multi-turn tasks

The future is open source. We need to embrace it and get on with it. Imagine if America closed the door on open source. We would explicitly be forcing American companies to pay $26-56 per 1MM tokens for the same intelligence their adversaries/competitors around the world would pay $0.50-1 for. This is economically unsustainable unless AI is bullshit. If it’s the supposed through line of all future economic activity that we’ve been told it is, we can’t handicap America at such a meaningful cost disadvantage. This can also be viewed thru a military lens and not just a simple economic one as well. Letting our adversaries attack us for $0.50 per 1MM tokens while we spend $26-56 per 1MM tokens to defend ourselves is equally ruinous. It’s the Cold War Soviet collapse in reverse.

we replicated anthropic jspace analysis on @thinkymachines Inkling new 1T model! it seems to be an outlier: where other models split into near-orthogonal sensory/workspace/motor blocks, inkling keeps roughly one geometry across the whole stack (early-late CKA ~0.8 vs ~0.5 elsewhere) we also look computed the J-space of @poolsideai's laguna XS 2.1 in bf16 vs nvfp4 to test the impact of quantization. result: almost none. the quantized model has the same jlens space as the non-quantized one both jspace checkpoint are up on hugging face!

Prime Intellect is built to provide the best price for any given task. if you're able to get better price/perf on any workload, would love to hear the details and look at it together — vincent@primeintellect.ai

Our team spent months developing RLFR, our method which uses probes on a model's internals as reward signals for RL. Silico reproduced it in 2 days, reducing hallucinations in Qwen3-8B by 37% without capability loss. (3/6)

or maybe let's make sure the most important technology in the history of humanity is not controled by just 4 men? Aka let's push for open science & open-source AI to distribute capabilities, power and wealth!

super-impressed by @PrimeIntellect 's work on this. we're not far off a huge explosion in rl with unsupervised environment generation, where devs build their own models by building their own envs. funnily enough, I've actually been banging on about roughly the same thing for a while now. back in the @argilla_io days we were convinced that datasets were the blocker to post-training. people just need an easy-to-use synthetic data generation library, plus a UI to review the dataset. due to cost, expertise, and algorithms, that wasn't really the case. but agentic rl feels very different now: - agents are better at helping people do post-training - small models are better, both relative to the frontier and in absolute terms. - environments teach actual tasks, not just emulation. - we have multiple working algortihms (GRPO, GKD, SDPO) that improve performance for common folk this time feels very different and the rest of the field needs to catchup or follow prime. anyway, just a personal back story on the rise of data in post-training. plugging harnesses into envs has been a huge focus of mine for the last 6 months, and I can see that taking it on with a trainer and env library makes way more sense. next steps for the community should be to unify some of these standard, whilst also maintaining our diversity.