According to NVIDIA's blog post, several companies took NVIDIA's open Nemotron models and post-trained them on proprietary data to build task-specific systems rather than relying on general-purpose closed models.
Across these three domains — computer use, law, and clinical dialogue — the common pattern in NVIDIA's account is the same: take an open Nemotron model and post-train it on domain-specific data rather than building or licensing a closed model from scratch.
NVIDIA's blog post frames cost reduction as a central argument for open models, citing three separate partner figures.
| Company | Base Model / Deployment | Reported Cost Difference | Benchmark Result |
|---|---|---|---|
| Harvey | Nemotron 3 Ultra (legal tasks) | At least 10x lower cost per run than leading closed models | Matches leading closed models on complex legal tasks |
| LangChain | Nemotron 3 Ultra (Deep Agents harness) | Approximately 10x lower cost per run than leading closed alternatives | Top agent accuracy among open models |
| Arcee AI | Nemotron on NVIDIA Blackwell | Roughly 90 cents per million output tokens — approximately 20x cheaper than comparable closed frontier models | Ranked second on PinchBench, fully open weight |
According to NVIDIA, LangChain tuned its Deep Agents framework for Nemotron 3 Ultra — adjusting prompts, tools, and middleware without retraining the model — and reported top agent accuracy among open models at roughly 10x lower cost per run than leading closed alternatives. (E3) NVIDIA also states that Arcee AI, running Nemotron on the Blackwell platform, achieved inference costs of about 90 cents per million output tokens, roughly 20x cheaper than comparable closed frontier models, while ranking second on PinchBench and keeping the model fully open weight. (E4) Taken together with Harvey's reported 10x figure, three independent partners each cite cost reductions clustered in the 10x-to-20x range against closed alternatives, according to NVIDIA's post.
Beyond single-task post-training, NVIDIA's blog highlights two applications where Nemotron is paired with agentic frameworks or larger models.
NVIDIA's post cites two healthcare companies building on Nemotron for clinical use cases.
Both claims are NVIDIA's characterization of partner outcomes; the blog post does not provide accuracy scores, error rates, or compute-cost figures for either company's clinical deployment.
NVIDIA's post also frames Nemotron's openness as enabling both language localization and cross-company collaboration.
All of the figures and quotes above come from a single NVIDIA blog post describing its own partners' results, so they should be read as NVIDIA's account of what those companies report — not independently verified benchmarks. Within that account, a consistent economic pattern emerges: three separate companies (Harvey, LangChain, Arcee AI) each report cost reductions in the same broad 10x-to-20x range relative to closed models, while claiming accuracy at or near parity — Harvey and LangChain by matching or topping open-model leaderboards, Arcee AI by ranking second on PinchBench. Alongside this cost pattern, the customization use cases NVIDIA cites span distinct domains — computer-use agents (H Company, 76% on OSWorld-Verified), enterprise search (Glean), clinical conversation and documentation (Abridge, Heidi Health), and language localization (YTL AI Labs) — suggesting that, according to NVIDIA's telling, the post-training approach is being applied across a wider range of specialized tasks than cost figures alone capture.
According to NVIDIA's blog post, Harvey post-trained Nemotron 3 Ultra on its legal benchmark and reached accuracy matching leading closed models on complex legal tasks at at least 10x lower cost per run.
NVIDIA's blog states Arcee AI ranked second on PinchBench while achieving inference costs of roughly 90 cents per million output tokens — about 20x cheaper than comparable closed frontier models — and keeping the model fully open weight.
According to NVIDIA, H Company post-trained Nemotron 3 Nano Omni into Holotron 3 Nano, which achieved higher than 76% accuracy on OSWorld-Verified, a benchmark for computer-use tasks.
According to Intel's newsroom announcement, the company is investing €5 billion ($5.7 billion) to expand its Leixlip campus in Ireland, producing Intel Xeon 6 and next-generation Xeon processors on the Intel 3 node. Intel says the capital programme began earlier this year and builds on more than €30 billion invested in Ireland since 1989.
According to TechCrunch and VentureBeat, Meta released Muse Code — a beta terminal-based AI coding agent for large repositories — alongside the Muse Spark 1.2 model on August 5, 2026. Muse Code installs with a single command, runs persistent background agents, and is entirely proprietary, putting Meta in direct competition with Anthropic's Claude Code and OpenAI's Codex.
NVIDIA has moved Alpamayo 2 Super into commercial use, according to NVIDIA's official blog. The autonomous-driving reasoning model runs on Cosmos 3 Super Reasoner, holds three times the parameters of its 10-billion-parameter predecessors, and ships under the OpenMDW-1.1 license, which ITHome reports NVIDIA has extended across the entire Alpamayo model family for fine-tuning and commercial redistribution.