Hermes 4.3 36B Centralized
Nous releases this centralized training run as a research artifact beside the Psyche-trained Hermes 4.3 model.
Research release
Training and model behavior
The card supplies a Torchtitan training configuration. It starts from Seed-OSS-36B-Base and describes a preset for 64 A100 GPUs. The training configuration specifies four epochs and a 131,072-token sequence length.Hermes 4.3 36B Centralized model card
Working with this model
Practical guidance. Use the control to compare training setups under the same evaluation. A training sequence limit alone does not establish the quality of long-context application results.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | NousResearch/Hermes-4.3-36B-centralized |
| Context window | 524,288 tokens Configured position limit; practical serving limits depend on the runtime. Hermes 4.3 36B Centralized configuration |
| Base model | ByteDance-Seed/Seed-OSS-36B-BaseHermes 4.3 36B Centralized model card |
Related models
- Hermes 4 405B
Nous Research · Hermes 4
A Llama 3.1 405B derivative with optional reasoning and structured output.
Launch date Downloadable - Hermes 4 14B
Nous Research · Hermes 4
A Qwen3-14B derivative with optional reasoning and structured output.
Launch date September 2025Downloadable - Hermes 4 70B
Nous Research · Hermes 4
A Llama 3.1 70B derivative with optional reasoning and structured output.
Launch date Downloadable