What are the hardware requirements for StableBeluga1 Delta?

StableBeluga1 Delta requires approximately 130.4 GB of VRAM and supports a context window of 2K tokens. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed StableBeluga1 Delta and how large is it?

StableBeluga1 Delta is developed by stabilityai, a model with 65b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

How does StableBeluga1 Delta perform on standard benchmarks?

StableBeluga1 Delta has the following published scores: MMLU 64.83. Compare against reference models on this page or on the LLM Explorer leaderboards.

StableBeluga1 Delta by stabilityai — VRAM 130.4GB, 2K context

Name: StableBeluga1 Delta
Author: stabilityai

StableBeluga1 Delta is an open-source language model by stabilityai. Features: 65b LLM, VRAM: 130.4GB, Context: 2K, License: cc-by-nc-4.0, LLM Explorer Score: 0.1, Arc: 68.2, HellaSwag: 85.9, MMLU: 64.8.

Arxiv:2302.13971 Arxiv:2306.02707 Dataset:conceptofmind/cot subm... Dataset:conceptofmind/flan2021... Dataset:conceptofmind/niv2 sub... Dataset:conceptofmind/t0 submi... En Endpoints compatible Llama Region:us Safetensors Sharded Tensorflow

Model Card on HF 🤗: https://huggingface.co/stabilityai/StableBeluga1-Delta

StableBeluga1 Delta Parameters and Internals

Model Type

auto-regressive, text generation

Additional Notes

Stable Beluga 1 cannot be used from the `stabilityai/StableBeluga1-Delta` weights alone. To obtain the correct model, the difference between LLaMA 65B and `stabilityai/StableBeluga1-Delta` must be applied using the provided script.

Supported Languages

English (Proficient)

Training Details

Data Sources:

conceptofmind/cot_submix_original, conceptofmind/flan2021_submix_original, conceptofmind/t0_submix_original, conceptofmind/niv2_submix_original

Methodology:

Supervised fine-tuning on internal Orca-style dataset

Model Architecture:

Stable Beluga 1 is an auto-regressive language model fine-tuned on LLaMA65B.

LLM Name	StableBeluga1 Delta
Repository 🤗	https://huggingface.co/stabilityai/StableBeluga1-Delta
Model Size	65b
Required VRAM	130.4 GB
Updated	2026-06-29
Maintainer	stabilityai
Model Type	llama
Model Files	9.9 GB: 1-of-14 9.7 GB: 2-of-14 9.7 GB: 3-of-14 9.7 GB: 4-of-14 9.7 GB: 5-of-14 9.7 GB: 6-of-14 9.7 GB: 7-of-14 9.7 GB: 8-of-14 9.7 GB: 9-of-14 9.7 GB: 10-of-14 9.7 GB: 11-of-14 9.7 GB: 12-of-14 9.7 GB: 13-of-14 4.1 GB: 14-of-14
Supported Languages	en
Model Architecture	LlamaForCausalLM
License	cc-by-nc-4.0
Context Length	2048
Model Max Length	2048
Transformers Version	4.32.0.dev0
Tokenizer Class	LlamaTokenizer
Beginning of Sentence Token	<s>
End of Sentence Token	</s>
Unk Token	<unk>
Vocabulary Size	32000
Torch Data Type	float16

Rank the StableBeluga1 Delta capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Best Alternatives to StableBeluga1 Delta

Best Alternatives	Context / RAM	Downloads	Likes
Kiyo 65M	2K / 0.1 GB	1456	2
Lite Oute 1 65M	2K / 0.3 GB	23	12
Lite Oute 1 65M Instruct	2K / 0.3 GB	15	12
Airoboros 65B Gpt4 M2.0	2K / 130.4 GB	850	0
Airoboros 65B Gpt4 2.0	2K / 130.4 GB	435	0
UltraLM 65B	2K / 130.4 GB	348	9
Openbuddy Llama 65B V8 Bf16	2K / 130.6 GB	177	9
Airoboros 65B Gpt4 1.4	2K / 130.4 GB	163	17
Llama 1 65B Hf	2K / 130.4 GB	14	1
Llama 65B Instruct	2K / 130.4 GB	198	15

Note: green Score (e.g. "73.2") means that the model is better than stabilityai/StableBeluga1-Delta.