CosmicAC Logo

Recommended model parameters

Recommended serving parameters, environment variables, and hardware for supported Managed Inference models.

CosmicAC recommends these serving parameters, environment variables, and hardware for the following models. The Job configuration reference defines each parameter. For how to apply them, see Create a Managed Inference Job.

Tensor parallel equals GPUs per replica for every vLLM model on this page. The web interface and the CLI take tensor parallel from the model master. If a job carries no tensor parallel value, CosmicAC uses GPUs per replica. Parakeet carries no tensor parallel value.

Qwen3-VL-235B-A22B-Thinking-FP8

Model ID Qwen/Qwen3-VL-235B-A22B-Thinking-FP8.

Serving parameters

ParameterValue
Runtime imagevllm/vllm-openai:v0.15.1
Data typeAuto
QuantisationNone
Tensor parallel8
GPUs per replica8
GPU memory utilization0.9
Max model length131072
Max concurrent sequences64
Reasoning parserdeepseek_r1
Video & image inputYes
Root disk size500 GB

Environment variables

TRUST_REMOTE_CODE=true
ENABLE_EXPERT_PARALLEL=true
ENFORCE_EAGER=true

Hardware

ResourceValue
GPUs8 H100 80 GB
CPU cores per GPU16
RAM per GPU150 GB

Qwen3.5-122B-A10B

Model ID Qwen/Qwen3.5-122B-A10B.

Serving parameters

ParameterValue
Runtime imagevllm/vllm-openai:v0.17.1
Data typeAuto
QuantisationFP8
Tensor parallel8
GPUs per replica8
GPU memory utilization0.9
Max model length32768
Max concurrent sequences32
Reasoning parserdeepseek_r1
Video & image inputYes
Root disk size500 GB

Environment variables

TRUST_REMOTE_CODE=true
SWAP_SPACE=0
ENABLE_EXPERT_PARALLEL=true
ENFORCE_EAGER=true

MiniMax M2.5

Model ID MiniMaxAI/MiniMax-M2.5.

Serving parameters

ParameterValue
Runtime imagevllm/vllm-openai:v0.15.1
Data typeAuto
QuantisationNone
Tensor parallel4
GPUs per replica4
GPU memory utilization0.85
Max model length27000
Max concurrent sequences256
Reasoning parserdeepseek_r1
Video & image inputYes
Root disk size500 GB

Environment variables

TRUST_REMOTE_CODE=true
SWAP_SPACE=0
ENABLE_EXPERT_PARALLEL=true
ENFORCE_EAGER=true

Hardware

ResourceValue
GPUs4 H100 80 GB
CPU cores per GPU16
RAM per GPU150 GB

Qwen2-VL-2B-Instruct

Model ID Qwen/Qwen2-VL-2B-Instruct.

Serving parameters

ParameterValue
Runtime imagevllm/vllm-openai:v0.15.1
Data typeAuto
QuantisationNone
Tensor parallel1
GPUs per replica1
GPU memory utilization0.9
Max model length32768
Max concurrent sequences64
Reasoning parserdefault, no parser
Video & image inputYes
Root disk size150 GB

Environment variables

TRUST_REMOTE_CODE=true
ENFORCE_EAGER=true
SWAP_SPACE=0

Parakeet

Model ID nvidia/parakeet-tdt-0.6b-v3. Parakeet is a speech-to-text model, so it takes no runtime image, data type, quantisation, tensor parallel, or reasoning parser. CosmicAC supplies the speech-to-text runtime image from the job type. The Parakeet Managed Inference parameters define each field, and Create a Parakeet Managed Inference Job shows how to apply them.

Serving parameters

ParameterValue
Chunk duration30 seconds
Chunk overlap5 seconds
Max file size1024 MB
GPUs per replica1
Root disk size250 GB
External disk size200 GB

Environment variables

MODEL_SOURCE=huggingface
HF_HOME=/mnt/external-disk/huggingface
NEMO_CACHE_DIR=/mnt/external-disk/nemo
HOST=0.0.0.0
INFERENCE_BACKEND=parakeet
PORT=8777
PARAKEET_STARTUP_TIMEOUT_MS=600000

Hardware

ResourceValue
GPUs1
CPU cores per GPU20
RAM per GPU125 GB

On this page