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.

h2oai logo

Can the RTX 5090 run h2o-lightning-31b?

h2o-lightning-31b·31.3B·GeForce RTX 5090·32 GB VRAM

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

Yes. The RTX 5090 runs h2o-lightning-31b fully in VRAM at Q6_K (about 28 GB of 32 GB), at roughly 38 tokens/s.

best quant
Q6_K
needs
28 GB
usable
32 GB
est. speed
~38 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the h2o-lightning-31b page →

Every quantisation of h2o-lightning-31b on the RTX 5090

QuantQualityMemoryFitSpeed
Q2_K Heavy loss13 GBfits~96 tok/s
Q3_K_M Noticeable loss17 GBfits~64 tok/s
Q4_K_M Small loss21 GBfits~52 tok/s
Q5_K_M Near-lossless24 GBfits~44 tok/s
Q6_K pickNear-lossless28 GBfits~38 tok/s
Q8_0 Lossless35 GBoffload~8.5 tok/s
FP16 / BF16 Lossless64 GBtoo big—

Speed is a bandwidth-based estimate for generating text (1792 GB/s); prompt processing and long contexts are slower.

Measured speed on the RTX 5090

No community reports for this pair yet.

Report your own speed →

Run h2o-lightning-31b on the RTX 5090

vLLM · Serve an OpenAI-compatible APIdocs
vllm serve h2oai/h2o-lightning-31b

Gated on Hugging Face: accept the licence on the model page and export HF_TOKEN first.

vLLM · Docker Compose (NVIDIA)docs
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: h2oai/h2o-lightning-31b
    ports:
      - "8000:8000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    environment:
      - HF_TOKEN=${HF_TOKEN}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 8000.

SGLang · Serve an OpenAI-compatible APIdocs
sglang serve h2oai/h2o-lightning-31b --host 0.0.0.0 --port 30000

Gated on Hugging Face: accept the licence on the model page and export HF_TOKEN first.

SGLang · Docker Compose (NVIDIA)docs
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve h2oai/h2o-lightning-31b --host 0.0.0.0 --port 30000
    ports:
      - "30000:30000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    environment:
      - HF_TOKEN=${HF_TOKEN}
    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="h2oai/h2o-lightning-31b", 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"])

Gated on Hugging Face: accept the licence on the model page and export HF_TOKEN first.

Vision models that fit the RTX 5090 comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
31.3BQ5_K_M≈29 GBFits in memory~44 tok/sCommercial OK
27.8BQ6_K≈25 GBFits in memory~43 tok/sCommercial OK
36BA3BQ6_K≈30 GBFits in memory~399 tok/sCommercial OK
meta-models logoMuse-Glimmer-30Bmeta-models
29.8BQ8_0≈30 GBFits in memory~31 tok/sCommercial OK
27.8BQ6_K≈25 GBFits in memory~43 tok/sCommercial OK
25.8BA4BQ8_0≈29 GBFits in memory~232 tok/sCommercial OK

Other GPUs that run h2o-lightning-31b

Frequently asked questions

Can the RTX 5090 run h2o-lightning-31b?
Yes. The RTX 5090 runs h2o-lightning-31b fully in VRAM at Q6_K (about 28 GB of 32 GB), at roughly 38 tokens/s.
How fast does h2o-lightning-31b run on the RTX 5090?
An estimate from memory bandwidth gives about 38 tokens/s at Q6_K.
Which quantisation of h2o-lightning-31b should I use on the RTX 5090?
Q6_K: the largest one that fits. Expected quality: near-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.