llmbenchio/benchmarks-by-vram

Updated on: 27 Sep 2026 Data contains: runs from the last 30 days Minimum runs: model/hardware combos with fewer than 3 runs are excluded

llm-bench.io — Community LLM Benchmark Leaderboard by Hardware

Per-model community benchmark data for local LLMs, curated from llm-bench.io and grouped by hardware and available VRAM.

This dataset contains only aggregated statistics derived from individual benchmark submissions. It does not contain raw submissions, prompts, model responses, machine identifiers, or client/session data. The raw data lives in the llm-bench.io database and is used here only to compute the summary stats below.

Why this dataset

LLM enthusiasts want to know: "How well does this model run on my hardware, and how good is the output?" This table gives a trustworthy, community-sized answer by filtering to model/hardware combinations with enough independent runs to be meaningful.

What's in it

Column	Description
vram_gb	Available VRAM (or unified memory) in GB on the tested hardware
hardware_kind	Platform category: Apple Silicon, NVIDIA, AMD
hardware	Specific device name (e.g. Apple M5 Max, NVIDIA GeForce RTX 4090)
model	Model name as run by the community (e.g. Qwen3.8-27B-oQ4e-mtp)
run_count	Number of independent community runs aggregated for this model/hardware
avg_quality	Mean quality score (0–100) across all four scenarios
avg_tok_s	Average generation speed in tokens/second
scenario_coding	Mean quality score for the code generation scenario (0–100)
scenario_agent	Mean quality score for long-horizon agent / task scenarios (0–100)
scenario_roleplay	Mean quality score for role-play / narrative scenarios (0–100)
scenario_research	Mean quality score for research & analysis scenarios (0–100)
scenario_*_runs	Number of runs feeding each scenario score
last_updated	ISO date of the most recent run contributing to this row
source	Link to the live site where results can be explored interactively

Methodology

Quality scores. Each submission includes scores on four tasks (code generation, long-horizon agent workflows, role-play, and research) that are scored by an automated LLM judge (a model used as an evaluator, not a human rater). Scores are 0–100.

Aggregation. For every (model, hardware) pair we:

1. Take all runs from the last 30 days.
2. Require at least 3 independent runs (this filters out noise and outliers — a single run is not trusted).
3. Average the per-scenario scores across those runs.

Hardware grouping. Rows are grouped by the device a model was tested on, so a 32 tok/s figure on a laptop and a 32 tok/s figure on a desktop are never mixed.

Privacy. The published table is a statistical summary. Individual submissions, prompts, model outputs, machine hashes, and client/session identifiers are never exported — they remain only in the llm-bench.io database.

How to use

import pandas as pd

df = pd.read_csv("benchmarks-by-vram.csv")

# Best coding model on an RTX 4090 with at least 4 runs
df[(df.hardware.str.contains("4090")) & (df.run_count >= 4)] \
    .sort_values("scenario_coding", ascending=False).head()

Or load directly in Python:

from datasets import load_dataset

ds = load_dataset("llmbenchio/benchmarks-by-vram")

License

Data is community-sourced and made available for reference and research use. Attribution is appreciated: source is llm-bench.io.