Prime Intellect

We're excited to join the NVIDIA Nemotron Coalition 💚 Frontier open models matter for the whole ecosystem. We're bringing the RL infrastructure and environments we've built over the last year to help scale agentic capabilities. Our contribution: the post‑training & RL environments layer. 2,500+ open RL environments, the verifiers framework, Prime Sandbox, and NeMo Gym integration — all natively integrated into NVIDIA's ecosystem, incl. Nemotron.

Reward hacking is the hardest problem in RL. We design settings where hacking is predictable, and find patterns between task difficulty and hack frequency. These runs are highly efficient, using <$1 in compute. We’re launching Sprints to allow everyone to join this effort. We share several of our findings here: https://lnkd.in/gkh7kEda

Today, we're sharing how our collaboration with NVIDIA helps power the open superintelligence stack. The next frontier of AI infrastructure is building systems for agentic models that can reason for hours, use tools, execute code, and learn from outcomes at scale. At Prime Intellect, we're building that stack end to end: → agentic RL training and inference on frontier open models → RL sandboxes → open-source libraries like verifiers + prime-rl Giving everyone access to frontier lab infrastructure. Our training and serving workloads run on NVIDIA Blackwell systems at rack scale and soon on NVIDIA Vera Rubin. This is powering the next generation of our training, inference, and sandbox stack. Inference is becoming central to both RL rollouts and production agents. We chose NVIDIA Dynamo because agentic inference at scale means handling global deployments, long-context reasoning, multi-turn trajectories, sparse MoEs, and large fleets of adapters. This collaboration also extends into research: → Nemotron models available through Lab → NVIDIA NeMo Gym environments in verifiers/prime-rl → NVIDIA NeMo RL stack integrated with verifiers Read more on our blog: https://lnkd.in/gsqhJvGa

Johannes Hagemann and Will Brown sat down with Sequoia Capital to explain how we're democratizing the frontier AI training stack. Allowing every company to turn their unique data into an agentic RL moat. https://lnkd.in/gCNypZ58

Introducing Lab: A full-stack platform for training your own agentic models Build, evaluate and train on your own environments at scale without managing the underlying infrastructure. Giving everyone their own frontier AI lab. https://lnkd.in/gqBjpqgn

We're excited to introduce Arcee AI's Trinity Large model. An open 400B total, 13B active parameter Mixture of Experts model with frontier performance. Trained in collaboration between Arcee AI, DatologyAI and Prime Intellect. Trinity Architecture Key design choices: - Interleaved local + global attention (3:1 pattern) - Grouped-query + gated attention - New load-balancing method (SMEBU) - Depth scaled sandwich norm and QK norm With extreme sparsity, built for long context and fast inference. Infrastructure - Large-scale synthetic data generation on 2k H100s - Training Trinity Large on 2k B300 GPUs Training stack: - Modified torchtitan - Muon optimizer - HSDP with FSDP group size 128 - Expert parallelism - Context parallelism for context extension - Improvements to recover quickly from hardware failures Benchmark Performance Despite lower active parameters, Trinity Large Base is very competitive with other top open models Inference Performance Trinity Large with its extreme sparsity and interleaved local/global attention results in: - Higher throughput - Lower end-to-end latency - Faster time to first token Next up: we’re continuing to scale agentic RL on top of Trinity Large using our frontier open post-training stack over the coming weeks. The full technical report by Arcee AI https://lnkd.in/db89eJdS The blog post by Arcee AI https://lnkd.in/g6NM5J3K

We believe the next breakthrough in long-horizon agents is training models to manage their own context. Introducing our new research direction on Recursive Language Models. We are sharing our initial experiments showing the promise of RLMs. https://lnkd.in/gQVidcK2

We're releasing Medmarks v0.1, the largest completely open-source automated evaluation suite for assessing the medical capabilities of LLMs! Developed in our MedARC community, with support from Prime Intellect So far we’ve explored 46 models to figure out the best! Why did we build this benchmark suite? Because there isn't yet a completely open and easy-to-run medical LLM benchmark, evaluated on various kinds of realistic tasks and updated regularly with new models. We aggregated a total of 28 tasks in 20 benchmarks in our suite. We divide them into two subsets: 1. Medmarks-Verifiable: 14 verifiable benchmarks, mostly multiple-choice question answering but also other tasks like medical calculations 2. Medmarks-OE: 6 open-ended benchmarks, e.g. answering patient questions On average, we find that GPT-5.1, GPT-5.2, and Qwen3-235B-A22B-Thinking are the best-performing models on the medical tasks evaluated. Lots of interesting observations: • medical-specific LLMs can be quite performant • open weight models are close to proprietary model perf but less token efficient • reasoning post-training improves performance • There's a few standout datasets that aren't saturated yet We built our evaluation suite on top of Prime Intellect's verifiers library. This provides us many advantages: 1. RL environments for free for datasets with training split 2. integration with Inference API enables easy benchmarking of model APIs 3. benchmarks available on Hub (under the MedARC organization) This is only just the beginning! We plan to add more models and benchmarks, and of course develop our own models and benchmarks! If you're interested in collaborating, join the MedARC Discord or contact us (contact@sophontai.com) To learn more about our benchmark suite, leaderboard, and interesting findings, read our blog post: https://lnkd.in/gNeZRVdw Check out the leaderboard: https://medmarks.ai code: https://lnkd.in/gNwmaXPZ environment hub: https://lnkd.in/gPCv3NUJ

Last week, we released INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more Blog Post: https://lnkd.in/gsgiaV_y Technical Report: https://lnkd.in/dVg8_8Af Chat: https://lnkd.in/gEi9acK6