"AI for people whose GPU budget is determined by what they can find on used online."
A structured database of real benchmark runs on budget hardware, with the methods, the scoring, and the raw records all public so anyone can check, reproduce, or extend it.
Most benchmark coverage assumes flagship hardware. Peasant Smith asks the question that actually matters to most people:
What can I really run on the hardware I can afford?
The honest answer almost always involves quantization. Full-precision
models are mostly inaccessible on consumer hardware - the useful, run-on-your-own-rig
results live at Q4, Q5, IQ3, and friends. That is why every record here
names its exact quant, and why runs with unknown quantization are labeled
unknown rather than guessed.
All August 2026 data below comes from a reference machine: two RTX 3060
12 GB, a Xeon E5-1650 v4, and 128 GB of DDR4 UDIMM desktop RAM (quad-channel,
X99 board). A second rig — three RTX 3060 12 GB, same CPU, 123 GB RAM (see
uranus-3x3060 in data/systems.csv) — contributed the
Ornith-1.5-35B-A3B and Qwen3.8-9B runs. t/s on the 3-card rig are not
directly comparable to the 2-card rows (extra GPU compute), so those two are
marked with †.
32 validated runs. Full table: LEADERBOARD.md.
| Model | Quant | Median t/s | Class | PS Points |
|---|---|---|---|---|
| LFM2.5-2.6B | unknown | 72.0 | excellent | 770 |
| Ornith-1.5-35B-A3B (MoE, 3B active) † | Q6_K | 48.8 | excellent | 558 |
| gemma4-26b-a4b (MoE, 4B active) | unknown | 44.1 | excellent | 491 |
| Qwen35-Hermes | unknown | 40.9 | excellent | 459 |
| Ornith-1.0-9B | Q6_K | 34.5 | excellent | 395 |
| Qwen3.8-9B † | Q8_0 | 32.2 | excellent | 438 |
| gemma-4-12b-it | UD-Q4_K_XL | 28.4 | excellent | 334 |
| Ling-3.0-Flash (MoE) | IQ4_XXS | 21.2 | excellent | 282 |
| Qwen3.8-27B | Q4_K_M | 18.9 @ 128k ctx | excellent | 269 |
| Muse-Glimmer-30B | UD-Q4_K_XL | 18.4 | excellent | 234 |
| gemma4-31b (dense) | unknown | 14.0 | good | 190 |
| Qwen3.5-122B-A10B-MTP (MoE) | UD-IQ3_XXS | 8.1 | good | 131 |
† = run on the separate 3× RTX 3060 rig (Uranus). Ornith-1.5-35B-A3B is also
covered in the deep-dive below; Qwen3.8-9B (empero-ai/Qwen3.8-9B, a Qwen3.5-9B-architecture
distillation of the 2.4T-A95B flagship) scored 10/11 on its fleet suite.
The pattern is the whole thesis: small efficient models and sparse MoE at Q4-class quants deliver interactive speeds on a few hundred dollars' worth of used GPUs; dense 30B+ needs patience. A 122B MoE at IQ3 still runs at 8 t/s with 128 GB RAM - a dense 31B at similar quality would not fit at all.
| You want... | Go to |
|---|---|
| "What runs on my GPU?" | data/benchmarks.csv + data/hardware.csv |
| How tests work & how they fail | docs/tests.md |
| The ranking formula | docs/scoring.md |
| Live rankings | LEADERBOARD.md |
| A tested system's full spec | data/systems.csv → hardware/systems/ |
| Run a benchmark yourself | scripts/benchmark/run_benchmark.py |
| Model-vs-model deep dives | docs/qwen27b-vs-ornith-15.md (dense 27B Q5 vs 35B-A3B MoE Q6, per-test results + reproduction) |
| Copy a full working setup | recipes/ — end-to-end recipes for specific model + hardware combos |
| Directory | Contents |
|---|---|
benchmarks/ |
Raw result records (JSON), schema, examples |
hardware/ |
System, GPU and CPU profiles (YAML) |
models/ |
Model profiles with quant guidance |
data/ |
Flat-file databases (CSV): benchmarks, hardware, systems, models |
docs/ |
Tests, scoring, methodology, definitions |
recipes/ |
End-to-end recipes: exact flags + results for a model on specific hardware |
scripts/ |
Benchmark runner, validators, scoring engine, migrations |
- Run the suite:
scripts/benchmark/run_benchmark.py <model> --hardware-ref <slug> - Validate:
scripts/validation/validate_benchmark.py <result.json> - PR the JSON into
benchmarks/raw/- failed runs included - Maintainers assign a
PS-NNNNid and regenerate the leaderboard
Details in CONTRIBUTING.md. Failed experiments are welcome - "won't fit on 12 GB" is the answer someone is searching for.
Every record preserves model + quant, backend and version, driver, OS, exact configuration, and operator. Screenshots alone are never accepted as data.
- Code: MIT · Data: CC BY 4.0 · Docs: CC BY-SA 4.0
- Methodology: docs/methodology.md · Schema: benchmarks/schema/benchmark-schema.json
Peasant Smith - because AI shouldn't require a flagship GPU budget.