Models

Predibase offers the ability to spin up private instances of nearly any open-source model available. These models fall into three categories:

1. Officially Supported LLMs: These are models we have first-class support, meaning they have been verified and are ensured to work well. They are also available as shared endpoints for non-VPC customers.
2. Best-Effort LLMs: These are models that have not been verified and may occasionally not deploy as expected.
3. Embedding Models (new ✨): Deploy a private instance of a variety of supported embedding models.

Officially Supported LLMs

Large Language Models (Text)

Deployment Name	Parameters	Architecture	License	Context Window*	Always On Shared Endpoint**
llama-3-3-70b	70 billion	Llama-3	Meta (request for commercial use)	8192	No
llama-3-2-1b	1 billion	Llama-3	Meta (request for commercial use)	32768	No
llama-3-2-1b-instruct	1 billion	Llama-3	Meta (request for commercial use)	32768	No
llama-3-2-3b	3 billion	Llama-3	Meta (request for commercial use)	32768	No
llama-3-2-3b-instruct	3 billion	Llama-3	Meta (request for commercial use)	32768	No
llama-3-1-8b	8 billion	Llama-3	Meta (request for commercial use)	63999	No
llama-3-1-8b-instruct	8 billion	Llama-3	Meta (request for commercial use)	63999	✅ Yes
llama-3-8b	8 billion	Llama-3	Meta (request for commercial use)	8192	No
llama-3-8b-instruct	8 billion	Llama-3	Meta (request for commercial use)	8192	No
llama-3-70b	70 billion	Llama-3	Meta (request for commercial use)	8192	No
llama-3-70b-instruct	70 billion	Llama-3	Meta (request for commercial use)	8192	No
llama-2-7b	7 billion	Llama-2	Meta (request for commercial use)	4096	No
llama-2-7b-chat	7 billion	Llama-2	Meta (request for commercial use)	4096	No
llama-2-13b	13 billion	Llama-2	Meta (request for commercial use)	4096	No
llama-2-13b-chat	13 billion	Llama-2	Meta (request for commercial use)	4096	No
llama-2-70b	70 billion	Llama-2	Meta (request for commercial use)	4096	No
llama-2-70b-chat	70 billion	Llama-2	Meta (request for commercial use)	4096	No
codellama-7b	7 billion	Llama-2	Meta (request for commercial use)	4096	No
codellama-7b-instruct	7 billion	Llama-2	Meta (request for commercial use)	4096	No
codellama-13b-instruct	13 billion	Llama-2	Meta (request for commercial use)	4096	No
codellama-70b-instruct	70 billion	Llama-2	Meta (request for commercial use)	4096	No
mistral-7b	7 billion	Mistral	Apache 2.0	32768	No
mistral-7b-instruct	7 billion	Mistral	Apache 2.0	32768	No
mistral-7b-instruct-v0-2	7 billion	Mistral	Apache 2.0	32768	✅ Yes
mistral-7b-instruct-v0-3	7 billion	Mistral	Apache 2.0	32768	No
mistral-nemo-12b-2407	12 billion	Mistral	Apache 2.0	65536	No
mistral-nemo-12b-instruct-2407	12 billion	Mistral	Apache 2.0	65536	No
mixtral-8x7b-v0-1	46.7 billion	Mixtral	Apache 2.0	32768	No
mixtral-8x7b-instruct-v0-1	46.7 billion	Mixtral	Apache 2.0	32768	No
solar-1-mini-chat-240612	10.7 billion	Llama	Custom License	32768	✅ Yes
solar-pro-preview-instruct-v2	22.1 billion	Solar	Custom License	4096	✅ Yes
solar-pro-241126	22.1 billion	Solar	Custom License	32768	✅ Yes
zephyr-7b-beta	7 billion	Mistral	MIT	8000	No
phi-2	2.7 billion	Phi-2	MIT	2048	No
phi-3-mini-4k-instruct	3.8 billion	Phi-3	MIT	4096	No
phi-3-5-mini-instruct	3.8 billion	Phi-3	MIT	65536	No
gemma-2b	2.5 billion	Gemma	Google	8192	No
gemma-2b-instruct	2.5 billion	Gemma	Google	8192	No
gemma-7b	8.5 billion	Gemma	Google	8192	No
gemma-7b-instruct	8.5 billion	Gemma	Google	8192	No
gemma-2-9b	9.24 billion	Gemma	Google	8192	No
gemma-2-9b-instruct	9.24 billion	Gemma	Google	8192	No
gemma-2-27b	27.2 billion	Gemma	Google	8192	No
gemma-2-27b-instruct	27.2 billion	Gemma	Google	8192	No
qwen2-5-1-5b	1.5 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-5-1-5b-instruct	1.5 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-5-7b	7 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-5-7b-instruct	7 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-5-14b	14 billion	Qwen	Tongyi Qianwen	32768	No
qwen2-5-14b-instruct	14 billion	Qwen	Tongyi Qianwen	32768	No
qwen2-5-32b	32 billion	Qwen	Tongyi Qianwen	16384	No
qwen2-5-32b-instruct	32 billion	Qwen	Tongyi Qianwen	16384	No
qwen2-1-5b	7.62 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-1-5b-instruct	7.62 billion	Qwen	Tongyi Qianwen	65536	No
qwen2-7b	7.62 billion	Qwen	Tongyi Qianwen	131072	No
qwen2-7b-instruct	7.62 billion	Qwen	Tongyi Qianwen	131072	No
qwen2-72b	72.7 billion	Qwen	Tongyi Qianwen	131072	No
qwen2-72b-instruct	72.7 billion	Qwen	Tongyi Qianwen	131072	No
solar-pro-preview-instruct (deprecated)	22.1 billion	Solar	Custom License	4096	No

*These context windows are well supported when using an A100 GPU. When using a smaller GPU, you may not be able to get the full context window.

**By default, all supported LLMs are availabled as shared endpoints. Models that are not "Always On" scale down to 0 and may have a brief spin up time before serving requests.

Visual Language Models

Deployment Name	Parameters	Architecture	License	Context Window*	Always On Shared Endpoint**
llama-3-2-11b-vision	10.7 billion	Mllama	Meta (request for commercial use)	32768	No
llama-3-2-11b-vision-instruct	10.7 billion	Mllama	Meta (request for commercial use)	32768	No

*These context windows are well supported when using an A100 GPU. When using a smaller GPU, you may not be able to get the full context window.

**By default, all supported VLMs are not available as shared endpoints. We will be adding support for VLMs as shared endpoints in the future. For now, you can created a private deployment to test base model or fine-tuned inference.

Best-effort LLMs

Predibase provides best-effort support for any Huggingface LLM meeting the following criteria:

• Uses one of the supported LoRAX architectures
• Has the "Text Generation" and "Transformer" tags
• Does not have a "custom_code" tag

How to Deploy a Custom LLM

1. Get the Huggingface ID for your model by clicking the the copy icon on the custom base model's page, ex. "BioMistral/BioMistral-7B".

2. Pass the Huggingface ID as the base_model, the appropriate accelerator ID for accelerator based on your tier or contract, and hf_token (your Huggingface token) if deploying a private model.

pb.deployments.create(
    name="my-biomistral-7b",
    config=DeploymentConfig(
        base_model="BioMistral/BioMistral-7B",
        accelerator="a100_80gb_100", # Required for custom models
        # hf_token="<YOUR HUGGINGFACE TOKEN>" # Required for private Huggingface models
        # cooldown_time=3600, # Value in seconds, defaults to 3600 (1hr)
        min_replicas=0, # Auto-scales to 0 replicas when not in use, set to 1 for an always-on deployment
        max_replicas=1
        custom_args=[ # Optional, see DeploymentConfig reference for further details
            "--max-input-length", "1000",
            "--max-batch-prefill-tokens", "100",
            "--max-total-tokens", "2048"
        ]
    )
)

caution

By default, we set safe defaults for --max-total-tokens for custom models, so if you'd like a larger context window per your model, be sure to set the --max-total-tokens in the custom_args.

3. Prompt your adapter as normal.

Instruction Templates

The following instruction templates are used in the UI when prompting our shared endpoints. When using the SDK or REST API for inference, you will need to include these templates yourself in the prompt, otherwise you may see less than stellar responses.

Llama 3 models

Instruct models

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a helpful, detailed, and polite artificial intelligence assistant. Your answers are clear and suitable for a professional environment.

If context is provided, answer using only the provided contextual information.<|eot_id|><|start_header_id|>user<|end_header_id|>

<insert your prompt here><|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n

Non-instruct models

None

Llama 2 models

Chat models

<<SYS>>
You are a helpful, detailed, and polite artificial intelligence assistant. Your answers are clear and suitable for a professional environment.

If context is provided, answer using only the provided contextual information.
<</SYS>>

[INST] <insert your prompt here> [/INST]

Non-chat models

None

Codellama models

codellama-13b-instruct

<s>[INST] <insert your prompt here> [/INST]

codellama-70b-instruct

<s>Source: user\n\n <insert your prompt here> <step> Source: assistant\nDestination: user\n\n

Mistral & Mixtral models

<<SYS>>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
<</SYS>>

[INST] <insert your prompt here> [/INST]

Solar models

Instruct models

<|im_start|>user\n <insert your prompt here> <|im_end|>\n<|im_start|>assistant\n

Non-instruct models

None.

Gemma models

Instruct models

<start_of_turn>user
<insert your prompt here><end_of_turn>
<start_of_turn>model

Non-instruct models

None

Phi-2

<|im_start|>user\n<insert your prompt here><|im_end|>\n

Zephyr-7b-beta

<|system|>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.</s>
<|user|>
<insert your prompt here></s>
<|assistant|>