--- title: "Prime Intellect Lets Teams Fine-Tune NVIDIA's Nemotron 3 Ultra Without a Single GPU" subtitle: "Prime Intellect Lab adds day-0 post-training for NVIDIA's 550B Nemotron 3 Ultra, opening the closed-lab agent loop to every team on Blackwell." url: "https://alphasignal.ai/news/prime-intellect-lets-teams-fine-tune-nvidia-s-nemotron-3-ultra-without-a-single" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Prime Intellect" company_url: "https://alphasignal.ai/company/primeintellect" published: "2026-06-04T19:48:32.000Z" updated: "2026-06-04T23:09:51.857Z" type: "News" topics: - "Llms" - "Agents" subtopics: - "Tool Use" - "Multi Agent" upvotes: 288 read_time: "5 min" source_url: "https://www.primeintellect.ai/blog/nemotron-3" comments: 0 --- # Prime Intellect Lets Teams Fine-Tune NVIDIA's Nemotron 3 Ultra Without a Single GPU > Prime Intellect Lab adds day-0 post-training for NVIDIA's 550B Nemotron 3 Ultra, opening the closed-lab agent loop to every team on Blackwell. Prime Intellect just turned NVIDIA's freshly released Nemotron 3 Ultra into something you can actually shape to your own workflow without standing up a single GPU. The hosted post-training platform, [Prime Intellect Lab](https://www.primeintellect.ai/blog/nemotron-3), now offers day-0 reinforcement learning, evaluations, and one-click inference for the new 550B open model, all running on Blackwell. The pitch is blunt: bring the task, skip the infrastructure. No cluster provisioning, no inference server configuration, no eval harness wiring. The same loop that closed labs run internally, environments, RL trainer, evals, deployment, continuous learning, is now exposed as a managed service against a frontier-grade open model. ## The model under the hood [Nemotron 3 Ultra](https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/) is not a conventional transformer. It is a 55B active, 550B total parameter Mixture-of-Experts hybrid Mamba-Transformer model that leverages Latent MoE, includes MTP Layers, and was pre-trained in NVFP4. The hybrid Mamba design is the reason it stays fast over very long agent traces, where pure transformers tend to grind to a halt as context balloons. The throughput claim is real: it achieves 5x higher throughput compared to other open models in its class, enabling long-running agents to complete tasks faster and more efficiently. On a pre-release endpoint, Nemotron 3 Ultra served over 300 tokens per second. On the benchmarks Prime Intellect cares about for agents, the model already sits at the open frontier out of the box: 95% on RULER at 1M context, 82% on IFBench, and SWEBench Verified scores between 65% and 70.4% across Pi, OpenHands, Hermes, OpenCode, and Mini SWE Agent, consistent performance regardless of which framework you deploy. By the [Artificial Analysis](https://artificialanalysis.ai/articles/nvidia-nemotron-3-ultra-launch-announced) Intelligence Index, Nemotron 3 Ultra scores 48, well ahead of the next strongest US open weights models, Gemma 4 31B (39), Nemotron 3 Super (36) and gpt-oss-120b (33), but behind the Chinese-led open weights frontier (Kimi K2.6 at 54). ## How NVIDIA actually trained it The training recipe matters because Prime Intellect is essentially handing you the same machinery to extend it. NVIDIA used Multi-Teacher On-Policy Distillation (MOPD): the student model generates rollouts across domains and receives dense reward signals from the corresponding teacher models. To maximize efficiency, MOPD runs asynchronously, with student rollout generation, teacher scoring, and student optimization fully pipelined. After producing an MOPD-trained checkpoint, new rounds of teacher training are initialized from the updated student model, and the improvements are merged into the next MOPD stage. This co-evolution between students and teachers enables continuous capability improvement and progressively stronger specialization across domains. The data story is unusually open. Building on a 10T token pre-training foundation, Nemotron 3 Ultra adds 212B new tokens targeting three high-value domain gaps: 4B tokens of synthetic legal data, increasing the proxy LegalBench average from 64.6% to 74.7%; 35B tokens of synthesized Wiki-based data, boosting proxy SimpleQA from 40.2% to 50.2%; 173B refreshed GitHub tokens through Sept. 30, 2025. This launch is also releasing 10M new SFT samples, 1M new RL tasks across multiple domains, and 15 net-new RL environments. ## What Day-0 on Lab actually gives you The Prime Intellect side of the integration covers the full loop: - Hosted RL training on Blackwell B200/B300 (and NVL72 where available) with native vLLM inference, NCCL weight broadcast, and the [Prime-RL](https://github.com/PrimeIntellect-ai/prime-rl) trainer. - 2,500+ open RL environments in the Environments Hub, spanning SWE, agentic search, tool-use, math, science, cybersecurity, and robotics sim, wrapped in the [verifiers](https://github.com/PrimeIntellect-ai/verifiers) framework. - Hosted evaluations on SWE-bench Verified, BrowseComp, tau-squared Bench Telecom, Terminal-Bench Hard, IFBench, AA-LCR, GPQA Diamond, and RULER-1M. - One-click inference with LoRA and full-weight adapters served via NVIDIA Dynamo for routing, autoscaling, and KV offloading. - Continuous improvement loops that funnel production traces back into the next RL run. The basic flow is three commands: ``` prime lab setup prime env install primeintellect/opencode-swe prime eval run opencode-swe -m nvidia/nemotron-3-ultra-550b-a55b --hosted ``` ## The numbers from the lab Prime Intellect has been ablating this stack on prior models for over a year, and the deltas suggest the approach transfers. Training INTELLECT-3.1 on the DeepDive environment lifted BrowseComp from 7.8% to 14.2%, nearly doubling performance on one of the least saturated agent benchmarks. SWE-bench Verified training on the R2E-Gym-Subset-Validated set produced monotonic gains for 30B-class MoE models. Two production case studies make the pattern concrete. [Ramp](https://www.primeintellect.ai/case-study/ramp) used Lab to post-train FastAsk, a 35B retrieval subagent based on Qwen3.5-35B-A3B, on its financial-spreadsheet workflow, reaching 66.25% exact-match accuracy against Claude Opus 4.6's 61.88%, roughly 27% faster than Opus, and gaining 10 points over its base model within hours of RL. [Zapier](https://www.primeintellect.ai/case-study/zapier) turned its AutomationBench from a static scorecard into a live RL training environment, catching reward hacking and fixing it in real time. ## Why this combination matters Frontier-grade open weights have existed for a while. What has not been broadly available is the infrastructure to actually specialize them with on-policy RL against real production traces. Nemotron 3 Ultra plus Prime Intellect Lab closes that gap on a single platform, on current-generation Blackwell silicon, against a model family explicitly post-trained for agent loops rather than chat. For teams building deep-research agents, coding copilots, or all-day enterprise automations, the calculus shifts. You can start from a model that already scores in the high 60s on SWE-bench Verified, point it at your environment, run hosted RL, and ship a specialized variant without owning the trainer, the inference layer, or the eval harness. That is the loop closed labs have had to themselves. It just got rented out. ## Sources - [X post by @primeintellect](https://x.com/primeintellect/status/2062622550300275088) - [Post-Training Nemotron 3 on Lab](https://www.primeintellect.ai/blog/nemotron-3) - [NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/) - [NVIDIA Nemotron 3 Ultra](https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra/) - [Nemotron 3 Ultra announced: high-speed, leading US open weights intelligence](https://artificialanalysis.ai/articles/nvidia-nemotron-3-ultra-launch-announced) - [Introducing NVIDIA Nemotron 3 Ultra: The Nemotron 3.x family is here!](https://www.baseten.co/blog/nvidia-nemotron-3x-model-family/)