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.

h2o-lightning-31b

Side-by-side specs, memory needs, licence, runtimes and benchmarks. Add up to 4 models.

--vram
any
--gpus
--quant
--add

On your GPU

  • h2o-lightning-31b
Property
Overview
TaskVision
Parameters31.3B
Active parametersdense
Context length—
Input → output
Released20 Oct 2026
Licenceh2o-model-non-commercial-licence-v1.1
Gated downloadyes, accept terms
Memory and speed
Fits your hardwarepick a GPU above≈21 GB
VRAM at Q4≈21 GB
VRAM at Q8≈35 GB
VRAM at FP16≈64 GB
Est. generation speedpick a GPU above—
Popularity
Downloads / 30 days690
Likes8
Runs with
llama.cpp
SGLang
Transformers
vLLM
Ollama tag—
Architecture
ArchitectureGemma4ForConditionalGeneration
Weights dtypebfloat16
GGUF availableno

Keep comparing

Memory figures are estimates for an 8K context at the chosen quantisation (Q4_K_M for LLMs by default). Speeds are rough, memory-bound estimates. Self-reported benchmark scores come from the model authors.