Thinking and instruct modes
According to the model card, the model follows a Qwen3-style ChatML template and switches modes by appending /think or /no_think to user input.
Open Source Model Profile · nvidia
Nemotron-Cascade-8B
Nemotron-Cascade-8B is an 8.19B-parameter Qwen3 text-generation model from nvidia. Its model card describes sequential reinforcement-learning post-training from Qwen3-8B-Base with thinking and instruct modes.
- Publisher
- nvidia
- Task
- text-generation
- Model type
- qwen3
- License
- nvidia-open-model-license
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Nemotron-Cascade-8B is published by nvidia as a Qwen3 text-generation model. The captured configuration identifies Qwen3ForCausalLM and Safetensors metadata reports 8,190,735,360 parameters. According to the model card, it is post-trained from Qwen3-8B-Base through multi-stage supervised fine-tuning followed by Cascade reinforcement learning, operating in thinking and instruct modes.
Recorded capabilities
Cascade RL pipeline
According to the model card, training begins with multi-stage supervised fine-tuning, followed by Cascade reinforcement learning across domains, with RLHF alignment described as a pre-step.
Qwen3-8B-Base post-training
According to the model card, Nemotron-Cascade-8B is post-trained from the Qwen3-8B-Base model.
NVIDIA Open Model License
According to the model card, use is governed by the NVIDIA Open Model License; hub card data records other.
Use cases in the source record
- General-purpose reasoning and instruction workflows using the card's documented thinking and non-reasoning instruct modes.
- Local deployment experiments following the card's sampling and YaRN RoPE-scaling recommendations.
Limitations and unknowns
- According to the model card, benchmark figures such as 84.0 MMLU, 75.5 MMLU Pro, 88.8 AIME 2024, and 74.5 LiveCodeBench v5 are publisher-reported Pass@1 scores, not independently verified measurements.
- Publisher superiority language in the model card reflects a publisher claim and was not independently verified as universal superiority.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: nvidia/Nemotron-Cascade-8B
Captured: Unknown. Processed: 2026-09-07T19:35:59.642581+00:00.
Nemotron-Cascade-8B Introduction We're excited to introduce Nemotron-Cascade-8B , a powerful general-purpose model trained through sequential and domain-wise reinforcement learning. Nemotron-Cascade-8B is post-trained from the Qwen3-8B-Base model. It operates in both thinking and instruct (non-reasoning) modes and delivers best-in-class performance across a wide range of benchmarks. Training Pipeline The training pipeline for Nemotron-Cascade begins with a multi-stage SFT phase to equip the model with foundational skills. Subsequently, Cascade RL is applied across multiple domains to further enhance the model’s performance in these…
F001F002F003F004F005F006F007F010F011F012F013F015F016F017