| Model | Score | Status | Released | Price in/out | Context | |
|---|---|---|---|---|---|---|
| Nemotron 3.5 Lightning fast open sub-agent for high-volume execution | 7.8 | GA | 2026-08-11 | — / — | 1.0M | Review → |
| Nemotron 3 Ultra self-hostable 1M-context reasoning | 8.5 | GA | 2026-06-04 | — / — | 1M | Review → |
Nemotron 3 Ultra is NVIDIA's frontier-adjacent open-weights flagship: a 550B-parameter hybrid MoE (55B active) released 2026-06-04 under the permissive OpenMDW-1.1 license, with a full 1M-token context and — unusually for mid-2026 — benchmarks you can actually reproduce, because the weights are public. It posts strong, verifiable coding and reasoning numbers (SWE-bench Verified 70.7, LiveCodeBench 89.0, MMLU-Pro 86.8) and a class-leading long-context result (RULER 94.7 at 1M tokens). It is text-only and multi-node to serve, but arriving just as Meta closed its open frontier line, it makes NVIDIA the credible Western open-weights anchor — and every self-hosted copy runs best on NVIDIA silicon.
Full review →Nemotron 3.5 Lightning is NVIDIA's small open-weights workhorse, released 2026-08-11 as the smallest member of the Nemotron 3 family: a 30B-parameter hybrid Mamba-2 / Transformer mixture-of-experts that activates only 3B parameters per token, published under OpenMDW-1.1 together with its training data and recipes. It is built for the execution layer of long-running agents (tool calls, result checks, formatting, sub-agent work), not for frontier reasoning, and it runs on a single H100, a DGX Spark or a GeForce RTX 5090. NVIDIA's own evaluation table is candid about where it sits: ahead of Nemotron 3 Nano and GPT-OSS 20B, but behind Qwen3.6-35B-A3B on 13 of 14 rows (SWE-bench Verified 51.56 vs 70.12). Pick it for speed, openness and cheap self-hosting on NVIDIA hardware, not for top-of-class accuracy.
Full review →