Rnj-1 Instruct
Release in the Rnj family · version rnj-1-instruct
Maintained by Essential AI1
The instruction-tuned version of Essential AI's Rnj-1, an 8.3B-parameter dense model trained from scratch. The base model was pretrained on 8.4T tokens at an 8K context, extended to a 32K context in a 380B-token mid-training stage, and then given a 150B-token supervised fine-tuning stage to produce this model.1
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Model hub: Rnj-1 base model card (external site: huggingface.co)
- Model hub: GGUF build for llama.cpp (external site: huggingface.co)
- Release notes: Rnj-1 announcement (external site: essential.ai)
- License: LICENSE (external site: huggingface.co)
Availability and license
Overall availability
Downloadable from Hugging Face without gating.1
Availability is separate from permission: read the license before using or redistributing.
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Base and instruction-tuned weights are both published, plus an official GGUF build.1 |
| Inference codeIs code for running the model published? | Public | The card documents use with Transformers 4.51.2 or later, vLLM (with the hermes tool-call parser), SGLang, and llama.cpp via the GGUF build.1 |
| Training codeIs the code used to train the model published? | Unknown | Not assessed. |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Unknown | The card says the model was trained on online web data and gives token counts per stage, but does not name the datasets.1 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The card lists the stage token budgets, context lengths, the Muon optimizer, the learning-rate schedule, and batch sizes; the fine-tuning data is not described.1 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | The card reports evaluation results, and Essential AI publishes the model's generations for its evaluations as Hugging Face datasets (for example rnj-1-instruct-evals).13 |
What it is useful for
The model card describes it for code generation across programming languages, agentic coding in frameworks such as mini-SWE-agent and Cline, tool calling, fill-in-the-middle code infilling, and math and science questions. It notes the model is mainly a coding and STEM model rather than one tuned for factual recall.1
Run and use notes
- The card recommends always using a system prompt and temperatures from 0 to 0.2, and warns that without them the model can truncate outputs or write code for non-code tasks. It documents extending the context to 128K with YaRN RoPE scaling by editing config.json and reports some regressions at that length, particularly on some science and performance evaluations.1
Organization context
Provenance and derivatives
Fine-tuned by Essential AI from its own Rnj-1 base model, which it trained from scratch. The card says the architecture is similar to Gemma 3 but uses only global attention, with YaRN for long-context extension.1
- Derived from: Rnj-1 (base) (external site: huggingface.co) — Essential AI base model; no separate catalog record.
Other releases in the Rnj family
- Rnj-1.5 InstructModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.