Frontier-beating quality at a fraction of the cost

With Plurai's vibe post-training, any engineer can turn an open-source model into a task-specific model that beats frontier LLMs, starting from a task description or a few examples. Plurai's platform runs the full pipeline, from synthetic data generation and post-training to benchmarking and inference optimization, on Nebius infrastructure built to train and serve at scale. For a global enterprise, we post-trained the NVIDIA Nemotron 3.5 Lightning open model to beat Claude Haiku-5.5 by more than 10 quality points at 44x lower cost, and to match OpenAI's GPT-6 Luna at 40x lower cost.

‍

The business challenge

Frontier models are the right tool for complex, open-ended tasks, but running every workload on them gets expensive fast. Many enterprise workloads are domain-specific and high-volume, such as extraction, classification and routing, and these are where a post-trained open model performs better at a fraction of the cost. Building one has meant labelled data, a data science team and weeks of iteration, so most teams keep everything on the frontier model.

‍

Vibe post-training: any engineer, any open model

Plurai removes the friction with vibe post-training. From a task description or a few examples, any engineer can post-train any open-source model, from NVIDIA Nemotron to Qwen and Llama, in about 20 minutes and get a model that beats frontier models on cost and accuracy. It works through Plurai’s web UI or directly in the IDE with the Plurai Claude Code plugin. The novelty underneath is grounded in Plurai’s published research: BARRED (ICML 2026), its method for generating synthetic training data for post-training, builds the training set instead of waiting for one.

Plurai also optimizes how its models run in production. Techniques such as Activated LoRA (aLoRA), token compression and smart decoding let many task-specific models share a single GPU without slowing down. In Plurai's published benchmark, aLoRA with shared prefix caching evaluated 100 adapters against a 5,000-token conversation in 285 ms on one GPU, 32x faster than standard LoRA.

Plurai makes post-training accessible to any engineer, and Nebius scales it across the open models it supports. For one of Plurai's enterprise design partners, Plurai post-trained two open models, NVIDIA Nemotron 3.5 Lightning and Qwen3.5, and both matched or beat the frontier models tested. Nemotron 3.5 Lightning delivered the biggest saving: 44x lower cost than Claude Haiku-5.5 with 10 points higher quality, and 40x lower than GPT-6 Luna at matching quality.

What the collaboration unlocks:

• For enterprises: high-volume workloads move off closed APIs onto open models they own, with better quality and a fraction of the cost, without hiring an ML team.
• For builders on Nebius: a path from a prompt to a post-trained open model, with post-training for supported models on Token Factory and serving on Nebius.
• For NVIDIA Nemotron: an open model released a few weeks ago, post-trained into a production-ready replacement for a closed model on a real enterprise workload.

‍

Results: higher quality at 44x lower cost

The post-trained Nemotron 3.5 Lightning model outscored Claude Haiku-5.5 by 10.4 points on real production chats (88.8 vs. 78.4) and 18.3 points on the held-out set (91.0 vs. 72.7), at 44x lower cost. It matched OpenAI's GPT-6 Luna on real production chats (88.8 vs. 88.6) and led it by 10.8 points on the held-out set (91.0 vs. 80.2), at 40x lower cost.
‍

‍

Figure 1. NVIDIA Nemotron 3.5 Lightning, post-trained by Plurai, matches GPT-6 Luna and beats the customer’s production model, GPT-4.1-mini, at 21x and 102x lower cost ($0.015 vs. $0.31 and $1.53 per 1,000 calls). Plurai post-trained two open models, Nemotron 3.5 Lightning (blue) and Qwen3.5-35B-A3B (orange); each trail runs from the out-of-the-box score to the post-trained score. GPT-6 Luna (with prompt caching), Claude Haiku-5.5 and GPT-4.1-mini are shown as-is. Quality: 0–100, average of 7 extracted fields on real production chats. Cost axis is log scale.

‍

Why Nebius: fine-tune and serve on one platform

Post-training is the step between a prototype and production, and it only pays off if the post-trained model can then be served at production scale. Nebius gives Plurai both on one platform.

• Fine-tune and serve in one place. Nebius Token Factory brings LoRA and full fine-tuning together with production inference, deploying a post-trained model straight to a dedicated endpoint, including NVFP4 serving on NVIDIA Blackwell.
• GPU capacity for custom work. Nebius AI Cloud supplies the compute for post-training and serving outside the managed path.
• Open-model choice. Token Factory post-trains 30+ open models, including Qwen, Llama, DeepSeek and GPT-OSS, so Plurai can pick the best base model for each task instead of locking into one.
• Production-grade serving. Dedicated endpoints with guaranteed performance and zero-retention inference in EU and US data centers, which enterprises need before moving a workload off a closed API.

‍

Why NVIDIA Nemotron 3.5 Lightning: the best accuracy per dollar

For high-volume workloads, the model that wins is the one with the best accuracy per dollar, and Nemotron 3.5 Lightning is built for exactly that tradeoff.

• Efficient by design. A 30B mixture-of-experts with only 3B parameters active per token, which NVIDIA positions as the highest-efficiency model in its class and as leading the accuracy-versus-speed frontier on the Artificial Analysis Intelligence Index.
• Built to be post-trained. Open weights that NVIDIA designed to be customized on an organization's own data, so post-training supplies the task-specific accuracy a small model lacks out of the box.
• Cheap to serve on NVIDIA hardware. It ships with an NVFP4 checkpoint optimized for NVIDIA Blackwell, which is what brought cost to $0.0078 per 1,000 calls in this case study.

In this case study, that tradeoff held: post-trained Lightning was both the cheapest system tested and the top scorer on the held-out set.

‍

How it works: from prompt to post-trained model in 20 minutes

Inside the case study, step by step

1. Describe the task. Plurai started from the customer's existing workload: reading each conversation and returning a structured set of business details.
2. Generate the data. Plurai built a training set of 3,722 conversations labelled by a frontier teacher model.
3. Post-train. Plurai fine-tuned NVIDIA Nemotron 3.5 Lightning and Qwen3.5-35B-A3B with LoRA on the same data.
4. Benchmark. Benchmark. Plurai scored both models against the customer’s production model, GPT-6 Luna and Claude Haiku-5.5, on real production conversations and a larger held-out set.
5. Optimize and serve. Plurai served Lightning in NVFP4 on an NVIDIA Blackwell GPU, bringing cost to $0.0078 per 1,000 calls with no change to the customer's pipeline.

‍

Get started

Plurai: Vibe post-train any open-source model for your task in about 20 minutes at app.plurai.ai, or straight from your IDE with the Plurai Claude Code plugin.

Nebius: Run post-training and inference for open models on Nebius Token Factory, and meet Plurai and Nebius at NVIDIA GTC Berlin, October 20–22, 2026.

‍

About Plurai

Plurai was founded by Dr. Ilan Kadar (CEO) and Dr. Elad Levi (CTO). Ilan, who holds a PhD in computer science and machine learning, is an AI leader with more than a decade of experience leading and building AI products, most recently as VP of AI at Nexar, where he led the company's AI efforts to build real-time, scalable and reliable systems for autonomous driving. Elad, who holds a PhD in mathematics, is an AI expert with more than a decade of experience developing innovative, high-impact AI products, most recently as Principal Scientist at Sightful and before that as AI Architect at Nexar. Together, they have published numerous research papers at top-tier machine learning conferences and released open-source projects with thousands of GitHub stars and hundreds of thousands of downloads.