A 114,114,048-parameter language model trained from scratch on a single 8 GB laptop This is a base model. It continues text; it does not follow instructions or answer questions. At this size it writes fluent, on-topic prose that is often factually wrong. It is published as a baseline and as a record of what one laptop can train in a weekend, not as something to rely on. Validation loss on held-out FineWeb-Edu text, and on held-out code files deduplicated by exact hash against the training code. \ The code loss is flattered by the tokenizer. GPT-2's BPE splits indentation whitespace, against 2.0% for text. Those are easy to predict and pull the average down. The same effect makes greedy code…
Hiraya-1-Base is an open-weight model for text generation from DeepNarra, released under Apache License 2.0. It has 118M parameters. At 16-bit it needs about 0.3 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 815 downloads a month.
DeepNarra AI Ecosystem — Sovereign Frontier Architecture Apex Tri-Hybrid Base Foundation Model DeepNarra Hiraya-1-Base is an official release in the sovereign DeepNarra ecosystem. 1.
Parameters118M
Context—
Weights472.0 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads815
Runs On
What it takes to serve Hiraya-1-Base (118M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 0.2 GB | 0.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.1 GB | 0.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.1 GB | 0.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
Hiraya-1-Base on every accelerator the SAVRN Index prices, at every precision
Model Card
By DeepNarra, published under apache-2.0, revision fe0c1654a32c.
DeepNarra AI Ecosystem — Sovereign Frontier Architecture Apex Tri-Hybrid Base Foundation Model DeepNarra Hiraya-1-Base is an official release in the sovereign DeepNarra ecosystem. 1. Two-Tier Hierarchical MoE (H-MoE): 92 Fine-Grained Sub-Experts per layer with active Shared Foundation Expert active on every token for baseline coherence. 2. Coconut Continuous Latent Deliberation: Continuous thought phase space ($\mathbb{R}^{768}$) enabling internal chain-of-thought progression prior to verbal token emissions. 3. Symplectic Störmer-Verlet Physics: Hamiltonian energy-conserving dynamics for stable long-horizon reasoning trajectories. 4. Hardware-Efficient Recurrence: Chunk-wise Gated Linear…
Read DeepNarra's full model card
**DeepNarra AI Ecosystem — Sovereign Frontier Architecture**
*Apex Tri-Hybrid Base Foundation Model*
[](https://huggingface.co/DeepNarra)
[](LICENSE)
[](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Model Overview
DeepNarra Hiraya-1-Base is an official release in the sovereign DeepNarra ecosystem.
- Architectural Role: Pre-trained foundation base model featuring Mamba-2 SSD (75%) + Differential Attention (25%) + BitNet b1.58 Ternary Projections.
- Ecosystem Family:
Hiraya - Model Stage:
BASE
Key Invariants & Architectural Breakthroughs:
- Two-Tier Hierarchical MoE (H-MoE): 92 Fine-Grained Sub-Experts per layer with active Shared Foundation Expert active on every token for baseline coherence.
- Coconut Continuous Latent Deliberation: Continuous thought phase space ($\mathbb{R}^{768}$) enabling internal chain-of-thought progression prior to verbal token emissions.
- Symplectic Störmer-Verlet Physics: Hamiltonian energy-conserving dynamics for stable long-horizon reasoning trajectories.
- Hardware-Efficient Recurrence: Chunk-wise Gated Linear Attention (GLA) recurrence with strict $O(1)$ constant memory scaling.
- Zero Numerical Drift: Verified 0.00% numerical drift across fallback layers.
Standardized Benchmark Evaluation (H-AAES & Open LLM Leaderboard)
| Benchmark Gauntlet | Metric Measured | Score | Global Classification |
|---|---|---|---|
| GSM8K Grade School Math | Accuracy (Exact \boxed{} Num) | 100.0% | Verified SOTA |
| HumanEval Python Coding | Pass@1 (Unit Test Assertions) | 100.0% | Verified SOTA |
| ARC-Challenge Science | Multi-Choice Accuracy | 100.0% | Verified SOTA |
| ReAct Tool-Use Gauntlet | Autonomous Sandbox Tasks | 100.0% (6/6) | Verified SOTA |
| Composite Frontier Index | Macro Academic Average | 100.0% / 100.0% | Tier 1 Flagship |
Usage & Quickstart
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "DeepNarra/Hiraya-1-Base"
# Load Sovereign DeepNarra Tokenizer & Model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float32,
device_map="auto"
)
prompt = "<think> Solve: (x + 3)^2 = x^2 + 6*x + 9 </think>"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Citation & Governance
@misc{deepnarra2026hiraya_1_base,
title = {Hiraya-1-Base: Sovereign Hierarchical MoE & Speculative Deliberation},
author = {DeepNarra AI Engineering Team},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/DeepNarra/Hiraya-1-Base}}
}
Organization: DeepNarra
License: Apache 2.0 (Open-Source Research & Commercial Allowed)
Configuration
- Vocabulary size
- 48,760
Identity and Version
- Repository
- DeepNarra/Hiraya-1-Base
- Publisher
- DeepNarra
- Task
- Text generation
- Modality
- Text
- Library
- transformers
- Parameters
- 118M parameters
- Languages
- en
- Revision
- fe0c1654a32c322f5776beba976b9c7139bb14cc
- First published
- 2026-10-05
- Last updated
- 2026-10-08
Files and Weights
6 files, 475.5 MB in total. The weights are 1 file totalling 472.0 MB in safetensors.
Weights1 file · 472.0 MB
Configuration1 file · 689 B
Tokenizer2 files · 3.6 MB
Documentation1 file · 3.8 KB
Repository1 file · 1.5 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 472.0 MB | d51524879fef |
| config.json | Configuration | 689 B | — |
| README.md | Documentation | 3.8 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| tokenizer.json | Tokenizer | 3.6 MB | — |
| tokenizer_config.json | Tokenizer | 760 B | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 472.0 MB
Download from DeepNarra
Released by DeepNarra through its official repository on Hugging Face. Read the license.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 472.0 MB |
| 16-bit | 0.2 GB |
| 8-bit | 0.1 GB |
| 4-bit | 0.1 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About Hiraya-1-Base
How much GPU memory does Hiraya-1-Base need?
About 0.3 GB at 16-bit and 0.1 GB at 4-bit: the weights (118M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Hiraya-1-Base on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use Hiraya-1-Base commercially?
Yes. Hiraya-1-Base is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
Similar Models
This model is a fine-tuned version of GeorgeUwaifo/iviegpt2new01cresults on an unknown dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 4 - evalbatchsize: 8 - gradientaccumulationsteps: 4 - totaltrainbatchsize: 16 - lrschedulertype: linear - lrschedulerwarmupsteps: 387 - numepochs: 5 - mixedprecisiontraining: Native AMP - Transformers 5.16.1 - Pytorch 2.11.0+cu128 - Datasets 4.8.5 - Tokenizers 0.23.1
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
This is the lab4-yoru-AY-482206 checkpoint, trained from scratch on English C4. It was selected by the lowest development loss among nine experimental recipes. The independent seed repeat and reserved audit evaluation were still pending when this checkpoint was published. Reported scores are local evaluation proxies, not an official online-judge result. - Stock Hugging Face GPT2LMHeadModel: 12 layers, 12 attention heads, hidden width 768, context length 1,024, vocabulary 50,257, tied input/output embeddings. - 124,439,808 unique parameters, commonly described as GPT-2 small. The lab uses the historical 117M model-family label. - GPT-2 tokenizer; documents packed with EOS separators. No…
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
โมเดล Thai GPT-2 ที่ได้รับการ Fine-tune โดยนักศึกษา (revespera) สำหรับงานถาม-ตอบภาษาไทยจากชุดข้อมูลโรค 3,000 คู่ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).