Nemotron-Nano-9B-v2
9B hybrid Mamba-Transformer with controllable thinking budget.
About
NVIDIA's 9B model built for fast reasoning: most attention layers are replaced with Mamba-2, so generating long thinking traces costs far less than a standard transformer.
A configurable thinking budget lets you trade accuracy against latency at inference time. NVIDIA released the weights, most of the training data, and the base checkpoints under Apache 2.0 — unusually generous for a frontier lab.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
On par with or better than Qwen3-8B on reasoning benchmarks with up to 6x higher inference throughput per NVIDIA's technical report.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Available free on OpenRouter and NVIDIA NIM; popular for single-GPU reasoning workloads; NVIDIA maintains the Nemotron line with regular releases; Apache 2.0 weights plus most training datasets and base checkpoints released. Manual review: hybrid Mamba-Transformer architecture and thinking-budget control verified in the technical report. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Available free on OpenRouter and NVIDIA NIM; popular for single-GPU reasoning workloads
- · NVIDIA maintains the Nemotron line with regular releases
- · Apache 2.0 weights plus most training datasets and base checkpoints released
- · manual review: hybrid Mamba-Transformer architecture and thinking-budget control verified in the technical report