selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

maritaca-ai logo

Can the RTX A4000 run S5-nano?

S5-nano·4.7B·RTX A4000·16 GB VRAM

Saved in this browser, used on every model page.
Yes

Yes. The RTX A4000 runs S5-nano fully in VRAM at FP16 / BF16 (about 10.0 GB of 16 GB), at roughly 26 tokens/s.

best quant
FP16 / BF16
needs
10.0 GB
usable
16 GB
est. speed
~26 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the S5-nano page →

Every quantisation of S5-nano on the RTX A4000

QuantQualityMemoryFitSpeed
Q2_K Heavy loss3.3 GBfits~161 tok/s
Q3_K_M Noticeable loss3.6 GBfits~108 tok/s
Q4_K_M Small loss4.2 GBfits~87 tok/s
Q5_K_M Near-lossless4.8 GBfits~74 tok/s
Q6_K Near-lossless5.2 GBfits~64 tok/s
Q8_0 Lossless6.0 GBfits~50 tok/s
FP16 / BF16 pickLossless10.0 GBfits~26 tok/s

Speed is a bandwidth-based estimate for generating text (448 GB/s); prompt processing and long contexts are slower.

Measured speed on the RTX A4000

No community reports for this pair yet.

Report your own speed →

Run S5-nano on the RTX A4000

Using FP16 / BF16, the best fit for RTX A4000.

Ollama · BF16 straight from Hugging Facedocs
ollama run hf.co/maritaca-ai/S5-nano-GGUF:BF16
llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf maritaca-ai/S5-nano-GGUF:BF16

Downloads the GGUF on first run. Installed through llama.app? The same flags work with `llama serve`.

LM Studio · Download with the lms CLIdocs
lms get maritaca-ai/S5-nano-GGUF@BF16

Then load it from the LM Studio app or with `lms load`.

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve maritaca-ai/S5-nano
vLLM · Smaller FP8 builddocs
vllm serve maritaca-ai/S5-nano-FP8
vLLM · Docker Compose (NVIDIA)docs
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: maritaca-ai/S5-nano
    ports:
      - "8000:8000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 8000.

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve maritaca-ai/S5-nano --host 0.0.0.0 --port 30000
SGLang · Docker Compose (NVIDIA)docs
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve maritaca-ai/S5-nano --host 0.0.0.0 --port 30000
    ports:
      - "30000:30000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 30000.

Transformers · Ask about an imagedocs
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="maritaca-ai/S5-nano", dtype="auto", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])

Vision models that fit the RTX A4000 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
4.7BQ8_0≈5.9 GBFits in memory~50 tok/sCommercial OK
4.3BQ8_0≈5.2 GBFits in memory~54 tok/sConditions
4.4BQ8_0≈5.8 GBFits in memory~52 tok/sCommercial OK
5BQ8_0≈5.6 GBFits in memory~47 tok/sConditions
4.3BQ8_0≈5.6 GBFits in memory~54 tok/sConditions
3.8BQ8_0≈4.0 GBFits in memory~62 tok/sNon-commercial

Other GPUs that run S5-nano

Frequently asked questions

Can the RTX A4000 run S5-nano?
Yes. The RTX A4000 runs S5-nano fully in VRAM at FP16 / BF16 (about 10.0 GB of 16 GB), at roughly 26 tokens/s.
How fast does S5-nano run on the RTX A4000?
An estimate from memory bandwidth gives about 26 tokens/s at FP16 / BF16.
Which quantisation of S5-nano should I use on the RTX A4000?
FP16 / BF16: the largest one that fits. Expected quality: lossless compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.