Nemotron-Labs TwoTower 30B-A3B Base
Nemotron-Labs TwoTower 30B-A3B Base is a released coding, long context, and classification model with open-weight and 128k context; evaluate it while provider pricing coverage matures.
Use it for
- Teams evaluating coding, long context, and classification
- Workloads that can use a 128k context window
Do not use it for
- Cost-sensitive launches that need sourced token pricing
- Vision or document-understanding workloads
- Strict JSON or tool-calling flows
- Family
- Nemotron-Labs TwoTower
- Released
- 2026-06-25
- Context
- 128k
- Max output
- 128,000
- Parameters
- ~60B total checkpoint; Hugging Face reports 63B params
- Architecture
- MoE + SSM Hybrid
- Knowledge cutoff
- 2025-06
- Specialization
- general
- Openness
- Open weights
- License
- NVIDIA Open ModelCommercial use: permitted
- Training
- Pretrained
No tracked provider token pricing is available yet.
About
Base text-generation checkpoint for Nemotron-Labs TwoTower. It uses a two-tower block-diffusion architecture over a Mamba-2/Transformer hybrid MoE backbone: a frozen causal AR/context tower processes clean prompt and committed tokens, while a trainable diffusion/denoiser tower fills token blocks by mask diffusion with cross-attention to the context tower. The checkpoint ships both towers and is not an instruction-tuned model.
Nemotron-Labs TwoTower 30B-A3B Base is an open-weight model in the Nemotron-Labs TwoTower family. The structured metadata tracks a 128k-token context window. Headline tracked benchmarks include Massive Multitask Language Understanding 78.2, MMLU PRO 60.9, and AI2 Reasoning Challenge 92.7.
Top use-case fit: coding, agents, and build tasks
Coding
1 relevant benchmark in the decision map.
Long context
Included by capability and metadata signals in the decision map.
Classification
2 relevant benchmarks in the decision map.
Provider price ladder
No tracked provider token pricing is available for this model yet.
Capabilities
No model capability flags are currently sourced.
Benchmark peer barsfor Coding
Benchmark scores(10)
| Benchmark | Score | Version | Source |
|---|---|---|---|
| Massive Multitask Language Understanding | 78.2 | 5-shot, accuracy; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| MMLU PRO | 60.9 | 5-shot, chain-of-thought exact match; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| AI2 Reasoning Challenge | 92.7 | ARC-Challenge, 25-shot, acc_norm; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| WinoGrande | 76.1 | 5-shot, accuracy; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| ReAding Comprehension Dataset From Examinations | 88.9 | 0-shot, accuracy; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| HumanEval | 75.6 | 0-shot; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| Mostly Basic Programming Problems | 74.3 | MBPP-Sanitized, 3-shot; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| Grade School Math 8K | 90.1 | 8-shot, accuracy; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| MATH-500 | 80.6 | 4-shot; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
| Multilingual Grade School Math | 80.4 | 8-shot, average accuracy; default TwoTower diffusion decoding at confidence_threshold=0.8, block_size=16, BF16 on 2xH100; evaluator/harness not published on the model card | https://huggingface.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16#benchmark-evaluations |
Migration checks
No linked migration route is available for this model yet.
API versions
v1.1v1.0Frequently asked questions
What is the context window of Nemotron-Labs TwoTower 30B-A3B Base?
Nemotron-Labs TwoTower 30B-A3B Base has a context window of 128k tokens.
What is the max output of Nemotron-Labs TwoTower 30B-A3B Base?
Nemotron-Labs TwoTower 30B-A3B Base can generate up to 128,000 output tokens.
When was Nemotron-Labs TwoTower 30B-A3B Base released?
Nemotron-Labs TwoTower 30B-A3B Base was released on 2026-06-25.
What benchmarks has Nemotron-Labs TwoTower 30B-A3B Base been tested on?
Nemotron-Labs TwoTower 30B-A3B Base has been evaluated on 10 benchmarks, including Massive Multitask Language Understanding, MMLU PRO, AI2 Reasoning Challenge, WinoGrande, ReAding Comprehension Dataset From Examinations.
No tracked provider token pricing is available yet.